Académique Documents
Professionnel Documents
Culture Documents
36
;;DOCUMI? B 412
DECISION-FEEDBACK EQUALIZATION
-W .& 04
FOR DIGITAL COMMUNICATION OVER DISPERSIVE CHANNELS
M. E. AUSTIN
LINCOLN LABORATORY
DECISION-FEEDBACK EQUALIZATION
FOR DIGITAL COMMUNICATION OVER DISPERSIVE CHANNELS
M. E. Austin
Abstract
Acknowledgments 84
References 85
iii
_____I_
___
DECISION- FEEDBACK EQUALIZATION
FOR DIGITAL COMMUNICATION OVER DISPERSIVE CHANNELS
Commercial and military demands for higher data rates in the transmission of digital
information over dispersive media, such as telephone links and the HF ionospheric channel, are
continually increasing. This has led to a widespread interest in the development of receivers
capable of mitigating the familiar intersymbol interference effects which inevitably accompany
such increased data rates. Receiver filters designed to partially compensate for the dispersive
characteristics of channels, or for the intersymbol interference distortion occurring when pass-
ing signals through dispersive channels, are referred to as" equalization" filters, or" equalizers."
The two principal objectives of this report are (a) to present a receiver structure which
uses its previous decisions to reduce intersymbol interference in high-speed data transmission
over dispersive channels - the decision-feedback equalizer, and (b) to compare the performance
of this new equalizer structure with that of the conventional equalizer structure currently being
employed.
In Sec. II, we first derive the optimal receiver for the problem of interest: synchronous
transmission of binary data over fixed, known dispersive channels. The main innovation here
is in the approach to receiver design in the presence of intersymbol interference - equalization
is attacked not as a problem in linear filtering theory as it has been previously, but rather from
a decision-theory viewpoint. This leads to the equations which specify the operations the re-
ceiver must perform on its input to minimize its probability of error. Our interpretation of
these equations results in the optimal equalizer structure, a nonlinear receiver which unfortu-
nately is too complex and impractical to construct, but which is nonetheless of interest for the
insight it affords us in considering sub-optimal equalization. By modifying the assumptions
used in arriving at the optimal equalizer, we next derive a sub-optimal equalizer structure
which we refer to as the "conventional" equalizer. We then determine the equations specifying
those parameter settings of the conventional equalizer which minimize the distortion energy
appearing at its output, and we prove the existence and uniqueness of their solution. Section II
concludes with a discussion relating this present work to that of earlier authors.
Adding to the assumptions used in deriving the conventional equalizer structure, we further
assume in Sec. III that the receiver has made no decision errors on previous bauds, leading to
our derivation of the decision-feedback equalizer. After interpreting the operation of this new
equalizer structure and the nature of its decision feedback, we optimize its parameters under a
minimum total output distortion energy criterion. Next, we consider conventional and decision-
feedback equalization with regard to the degrees of freedom associated with their respective
structures, and their manner of treating the intersymbol interference arising from both past
and future bauds. We conclude Sec. III with heuristic arguments for the error-propagation ef-
fects anticipated in the operation of the decision-feedback equalizer.
With a view to obtaining a quantitative comparison of the performances of the conventional
and decision-feedback equalizers, in Sec. IV we begin by deriving the channels we wish to study -
the maximal distortion channels. For channels with a given overall channel dispersion length,
the maximal distortion channels exhibit the largest possible distortion under a sum of the side-
lobe magnitudes distortion measure. We then seek a method of evaluating the performance of
the conventional equalizer. After noting that a direct calculation of the performance is concep-
tually possible but computationally impractical, we develop an alternative approach. A new
computational algorithm is derived for obtaining upper and lower bounds on the performance of
the conventional equalizer to any specified degree of accuracy. The implementation of this al-
gorithm is explained, and its efficiency is studied through its application to examples of interest.
Next, we present the exact performance data on the conventional equalizer when it is applied to
equalization of the maximal distortion channels, as determined using this new algorithm. Sec-
tion IV concludes with a discussion of the effects of equalizer length on the performance, and of
the trade-off possible between signal-to-noise ratio (SNR) and equalizer length at different levels
of performance.
In Sec. V, we are concerned with determining the performance characteristics of the decision-
feedback equalizer. Next to its overall error rates as a function of SNR for various channels,
error-propagation effects are of prime concern. Because of the nonlinear nature of the decision-
feedback equalizer structure, its performance was studied through digital simulations at low and
intermediate SNR. However, at high SNR, errors occur too infrequently to determine the error
rates and error-propagation behavior accurately through simulations of reasonable duration.
Thus, we determine the necessary modifications to our earlier algorithm to develop a second
algorithm - a mixture of computation and simulation - which overcomes this difficulty, enabling
us to obtain the desired error rates as well as burst and guard-space data, even at high SNR.
We next present the results of our performance studies at all SNR, with discussions of overall
error rates, burst and guard-space data, and the effects of equalizer length on these quantities.
Finally, we conclude Sec. V with a development of an efficient method of bounding the mean burst
duration, by modeling error propagation as a discrete Markov chain process, enabling us to
treat equalizer recovery from the burst-error mode as a first-passage-time problem. This
technique, which we find to be applicable at intermediate SNR due to the correlation properties
of the distortion appearing at the output of the decision-feedback equalizer, provides us with a
method of obtaining bounds on the mean recovery time without having to resort to simulations.
Section VI includes our conclusions, and suggestions for further research. After discuss-
ing threshold effects noted in the operation of the decision-feedback equalizer at low SNR, and
its obvious advantages over the conventional equalizer at high SNR, we then compare both equal-
izer performances to upper bounds to obtain some feeling for the degree to which each is sub-
optimal. Differences in the performances of the two equalizers are then interpreted, and ex-
plained in terms of their respective methods of treating intersymbol interference in general,
and in terms of their residual noise and intersymbol interference distortions, particularly at
high SNR. Based upon our error-burst and guard-space studies, we note the additional im-
provements possible in the decision-feedback equalizer operation which one might realize through
application of burst-error detecting and correcting coding schemes. Finally, we discuss briefly
some issues which may be of interest in further research, such as extensions of the present
work to handle colored noise, correlated message sequences, multilevel signaling schemes,
__
different types of decision-feedback data, and the sensitivity properties of the conventional and
decision-feedback equalizers with respect to channel measurement errors.
Having outlined the purpose and content of this report, we will devote the remainder of Sec. I
to presenting some background material of interest. First, we introduce an important device
for implementing equalizers which has been employed in all equalization systems of recent de-
velopment - the transversal filter; next, we point out the difference between channel equaliza-
4
tion and equalization of intersymbol interference, and the desirability of the latter; finally, we
present a classification of equalizers and brief descriptions of three adaptive equalizers cur-
rently under development. Two of these adaptive equalizers employ decision feedback, but both
for a different purpose than the decision-feedback equalizer presented in this report.
Early efforts at channel equalization date back to 1928, when Zobel1 published an extensive
work on distortion correction using lumped RLC filters. For some time, such filters adequately
provided the amplitude compensation desired for telephone circuits, for the ear proved rather
insensitive to the phase distortions they generally introduced. With the development of televi-
sion, and with digital communication aiming at increased data rates, phase distortion became
of greater concern, leading to the work with the transversal filter reported by Kallmann 2 in
1940. The analog transversal filter may be realized using a tapped-delay-line (TDL) terminated
in its characteristic impedance, with a high-input-impedance amplifier at each tap,output. With
the amplifier gains suitably chosen for the desired equalization properties, the amplifier outputs
are summed to obtain the transversal filter output, as indicated in Fig. 1(a). Often, the TDL
1ul
(b)
Fig. 1. Transversal filter. (a) Block diagram of filter; (b) representation in this report
of both analog and digital realizations.
____ ·_ _ ___
input is sampled, the delay lines are replaced by digital shift registers, and the weighting and
summing operations are performed digitally. Throughout this report, whether considering an-
alog or digital realizations, we will draw transversal filters as shown in Fig. 1(b). The tap gains
of the transversal filter can be set to render amplitude compensation without phase distortion,
phase compensation without amplitude distortion, or both amplitude and phase compensation.
Such versatility, coupled with the increasing speed and availability of computers for calculating
and adjusting its gains, has caused the digital transversal filter to become increasingly impor-
tant in recent years, as illustrated in the adaptive equalization systems described further below.
One of the first adaptive receivers designed to equalize a dispersive channel was the RAKE
system described by Price and Green. 3 This system transmits wide-band message waveforms
which are correlated against their delayed replicas at the receiver to enable measurement of
the amplitude and phase characteristics of the individual paths of a multipath medium such as
the HF ionospheric channel. By use of these measurements to set the amplitude and phase of
the tap gains of its transversal filter, the RAKE receiver achieves nearly optimum weighting of
the waveforms received from each of the paths, and sums them coherently to form its output.
Considering the low 45-bits/sec data rate it achieved on the HF ionospheric channel, the RAKE
system represents a highly inefficient use of its 10-kHz bandwidth when compared, for example,
with the less sophisticated Kineplex system which was designed for data rates of up to 3000bits/sec
5
with a bandwidth of only about 3.4 kHz, and without attempting channel measurement. This
serves to illustrate that, for data transmission purposes, one should not necessarily strive to
equalize the channel itself, but rather should consider equalization for a particular choice of
signals to be transmitted over the channel, as pointed out in the following example.
It is easily shown through the z-transform approach to inverse filtering that, in the absence
of significant noise, the discrete two-path channel of Fig. 2(a) can be equalized using the trans-
versal filter of Fig. 1(b), if its taps are spaced by the time delay between the two paths T, and
1 13-42-10140J
I"~~~~~
- ______ - __ ___ __ I~~~
rT
(a)
bT
I
t
O T
(C)
4
its gains are set according to the relationship a n = (-b)n - i
(see Ref. 6). If this equalizer is
truncated after N taps, there will be residual response having magnitude Ibl N On the other
hand, suppose we plan instead to use the channel of Fig. 2(a) to transmit data using amplitude
modulation of the signal waveform shown in Fig. 2(b), which has a baud duration T > . Syn-
chronous sampling of the channel output as indicated in Fig. 2(c) renders a sampled waveform
which may be equalized with a transversal filter with gains an = [-(bT/T)]n - 1 and with tap spac-
ing T. With this latter approach, it is seen that convergence of the tap gains is obtained more
rapidly than before. Moreover, the tap spacings are independent of -, an important matter if
T is a time-varying quantity, for then the channel equalization receiver must employ many more
taps than are required by the intersymbol interference equalizer, which need only adjust its tap
gains. This simple example points out that if one intends to transmit a signal s(t) over a dis-
persive channel having an impulse response g(t), it is generally more efficient to minimize the
intersymbol interference distortion appearing at the receiver output through equalization of the
equivalent impulse response h(t) = s(t) ( g(t), rather than attempting to equalize the channel
g(t) itself in order to be able to handle arbitrary s(t).
Equalization filters fall within three categories: fixed, automatic, and adaptive. Fixed
equalizers are adjusted to provide the amplitude and phase compensation necessary to correct
the average distortion characteristics of dispersive channels, while automatic and adaptive
equalizers, on the other hand, have generally found application in synchronous data transmission
to equalize the equivalent channel, i.e., reducing the intersymbol interference distortion at the
sampling times. The automatic equalizers transmit their known pulse waveforms prior to data
transmission, which they utilize in adjusting the tap gains of their transversal equalization fil-
ter (for example, see Ref. 7). Adaptive equalizers differ from the automatic equalizers in that
they adjust their transversal filter parameters during data transmission, using either a sound-
ing waveform or the message waveform itself to minimize the intersymbol interference distor-
tion, thus indirectly obtaining equivalent channel measurement. Three such adaptive systems
are described briefly below.
An adaptive equalizer named ADAPTICOM has been under study at Cardion Electronics, Inc.
and has been reported by DiToro. 8 This system uses approximately a 3-kHz bandwidth to achieve
a degree of equalization by using a filter which is matched to the received waveform for a single
pulse transmission, in cascade with a transversal filter. ADAPTICOM's parameter settings
are determined by periodically interrupting the data transmission several times per correlation
time of the channel and sounding the channel, although apparently no attempt is made to exploit
the correlation between the successive measurements to improve the accuracy of the resultant
settings. The transversal filter centertap has unity gain, with the remaining tap gains set equal
to their respective tap outputs, when a single received pulse is passed through the matched fil-
ter, at that instant when the centertap output attains its maximum value. Thus, ideally, the tap
gains are set equal to the sampled autocorrelation function of the equivalent channel. DiToro
states that such gains, together with proper adjustment of a single gain simultaneously affecting
the total contribution of all but the centertap outputs, minimize the mean square distortion due
to intersymbol interference at the output, although his mathematical basis for this has not yet
been published. Performing satisfactorily in the laboratory using a multipath simulator,
ADAPTICOM is being constructed for testing on an actual HF ionospheric link.
Lucky 9 reports the development and testing of an adaptive equalizer using a digital trans-
versal filter, for application to telephone-line equalization. Similar in other respects to
_ __
II. OPTIMAL AND SUB-OPTIMAL EQUALIZATION
Here, we consider the problem of determining the structures of receivers for digital
communication over a linear dispersive channel. As shown in the model of Fig. 3, during the
kth -baud period of duration T, ks(t) is transmitted, where s(t) vanishes outside an interval of
t width T and where ~k = +1 or -1, corresponding to hypotheses Hi and H0 , respectively. We
assume that the equivalent channel impulse response, h(t) = s(t) ( g(t), is known to the receiver
through prior measurement made either directly or indirectly, as in the case of the adaptive
equalizers described in Sec.I. In addition, the channel has additive noise n(t) corrupting its out-
put. Thus, the input to our receiver is
where here, and throughout the remainder of the report, Z denotes the summation for k = -o
to +0 unless otherwise noted. k
n( r(t)
'- - s(t) g(t) + r(t)
SIGNAL ACTUAL
FILTER CHANNEL
(k-1
(a)
n(t)
l. l
Xth(t) + r(t)
EQUIVALENT
k-1 CHANNEL
(b) 13-42-0141
Given the received waveform Eq. (1) for all t, our problem is to obtain receivers for making
decisions on the k.' We first derive the optimal receiver structure, and then a simpler sub-
optimal receiver. We will refer to the latter receiver as the "conventional" equalizer, to dis-
tinguish it from the "decision-feedback" equalizer to be presented in Sec. III.
In deriving the optimal receiver for the problem stated above, we make the following
assumptions:
These assumptions are generally met in practice. The extension to colored Gaussian noise is
straightforward, as discussed in Sec. VI. Under these assumptions, it is a well-known result
7
in statistical decision theory that, to make a decision on the zeroth baud as to the value of 0,
the optimal receiver computes the likelihood ratio
p[r(t)l0 = +1]
(2)
p[r(t) | 0 = -1]
p[r(t) 0] = E E p[r(t)l
o -0,
(t] P[, _+] (5)
+-
where the summations are over all possible message sequences occurring before and after the
zeroth baud, respectively. From Eq. (1) and Assumption (3), we may write
2
where we have defined K1 = K exp [-(1/NO ) f r (t) dt], and the quantities
a k A NS
2 r(t) h(t - kT) dt= r(t + kT) h(t) dt (7)
and
Note that all the information of the received waveform has been condensed into the set of suf-
ficient statistics {ak), which may be generated as shown in Fig. 4. In Fig. 4(a), we see that the
received waveform is input to a TDL, thus simultaneously making available the shifted versions
of r(t) required in Eq. (7) where it is indicated that each tap output is to be multiplied by [(2/N 0 ) x
h(t)] and the product integrated over all time t. Each integrator output will remain constant,
however, after h(t) becomes zero (or negligibly small). Thus, if h(t) is nonzero only over the
8
2
No
an O-n
a0
(a)
SAMPLE AT
BAUD RATE
h(-t) TDL
MATCHED .. I
FILTER
a n o 0 a n
(b)
interval (0, T 1 ), then we may sample each integrator output at any time t > T to obtain the de-
sired ak . Note that if we sample at t = T 1 , then we may replace the multiplication-integration
operation on each tap output by a linear filter having an impulse response [(2/N 0 ) h(-t)], since
the output of such filters will also render the desired ak at t = T 1. Moreover, since the linear
filters common to each tap may be replaced by a single filter operating on the received waveform
r(t) before it enters the TDL, and since the simultaneous sampling of each tap output may clearly
be replaced by the sequential sampling of the waveform into the TDL at the baud rate 1/T, then
an equivalent method of generating the sufficient statistics is as shown in Fig. 4(b). Thus, we
note that optimal reception involves passing the received waveform through a filter matched
(except for the gain of 2/N 0 ) to the equivalent channel impulse response h(t), sampling the matched
filter output at the baud rate to obtain the set of sufficient statistics {ak}, and then passing these
into a TDL. Before considering the additional operations which the optimal receiver performs
on these sufficient statistics, we first consider the bk sequence defined by Eq. (8).
From a comparison of Eqs. (7) and (8) and the discussion above, we observe that the bk se-
quence may be generated using a structure identical (except for a gain of 2) to that of Fig. 4(b), if
the input r(t) is replaced by h(t). But this corresponds to sampling the matched filter response to
a single pulse transmission in the absence of noise, thus rendering (except for the gain of 1/No)
the samples of the equivalent channel autocorrelation function, and bk = b k as indicated in
Fig. 5. Returning to Eqs. (5) and (6), we may write
9
SAMPLE AT
BAUDRATE
TDL
T_~~~~~~~
bn
I I I -.- I I
b1 bo b_! b-n
! ! I rr
b-4 I b4
M
b-3 b
$
p[r(t)Io] = K2 exp[ oa o0 -
2 i2bo] 2
+
exp[ 4
kak j k (j kbj-k] (10)
i 1 k#0 j k¢=j
4
Using this expression in Eq. (2), and canceling factors common to the numerator and denominator,
we find the likelihood ratio to be
exp[ 4
kak- k 4jkbj-k1 I
z- ki j ksi j bj- k =1
44~
A = exp[2a 0 ] (11)
exp z 4
kak -
Z z 4j kbi-,ki
+- Ik7& 0 j k:/i
44~ 5,=-j
If we denote
t--, ( ·, ~+)
0,
1+ A g1
4o=
and
-_ 1 (12)
0o~
4 (13)
C(5) = 23 jkbj-k
j kj
then, noting that the optimal receiver uses the decision rule
10
>0 decide Hi
in A
<0 decide H0 (14)
I r) 13-42-107441
As shown above, the optimal equalizer structure passes the received waveform through a
matched filter, sampler, and TDL to obtain the set of sufficient statistics {ak}. The sufficient
statistic derived from the interval upon which the decision is being made, a 0 , is connected
directly to the output through an amplifier of gain two. The statistics derived from all other
intervals are weighted by all possible 4- and + sequences, as indicated. For any particular
sequence, bias terms, the C(s), are subtracted from the weighted sum of the ak, and the differ-
ences passed through exponentiators before summing with similar terms from all other com-
binations of - and 4 . The natural logarithms of these overall sums are then added and sub-
tracted respectively to Za 0 , as shown in Fig. 6. This last summation result is compared with a
zero threshold to decide between the hypotheses H i and H0.
Next, we want to make a few observations concerning this optimal equalizer structure. First
we note that if C(s+) = C(0) = 0 for all choices of 4_ and , then the inputs to their respective
summations from the exponentiators will be identical, hence the logarithms of these sums will
cancel with each other at the overall output. When this occurs, the decision statistic is simply
2a 0 , the other ak do not appear, and the entire receiver degenerates to a matched filter.
11
The complexity of the optimal equalizer structure derived in Sec. A stems from the large
number of terms entering into the calculation of the numerator and denominator of the likelihood
ratio, a term being required for every possible message sequence and 4+. To simplify cal-
culation of the likelihood ratio and obtain a less complex receiver structure, we make an addi-
tional assumption similar to that which proved fruitful in a radar resolution problem considered
in Ref. 12, namely,
(4) k are Gaussian random variables of zero mean and unit variance
for k 5 0.
In making this assumption, of course, we have deviated from the digital communication problem
of interest, the assumption being perhaps more appropriate in a PAM communication system
context. Nonetheless, we will see that this assumption enables us to arrive at a much simpler,
though sub-optimal, equalizer than the optimal equalizer derived in Sec. A. Moreover, we employ
the assumption only in arriving at the equalizer structure and, having obtained the structure, we
will later optimize its parameters for the actual digital communication problem at hand. Using
this assumption, then, in addition to our earlier assumptions, the equation corresponding to
Eq. (9) above becomes
12
where, from Assumptions (1) and (4), it follows that
P[t-, +] = K 4 exp -
kO
k=/ 0 k2
1 2 + t1 2
= K 4 exp [-2 k 2
k
where K5 A K e /
and 6jk is the Kronecker delta, zero for j # k and unity for j = k. From
Eqs. (6) and (16), the integrand of Eq. (15) becomes
p[r(t) 4 0 ] . p[,
]p[ +] =
KK exp
k
4
kak -
j
, Z
k
j(bjk + 2 2jjk) k kk
15kk 6 (17)
To simplify notation, it becomes convenient at this point to introduce column vectors x, a, and
c defined by
a c
-2 '-2 -2
a ,-I C-1
1_
x A a c A (18)
a1 ct
a2 c2
where we define
Ck bk + 6k (19)
and the ak and k are as defined previously. We also introduce a matrix Q defined by its elements
1 (20)
Qjk bj-k + 2 jk (0)
for all j, k # 0. With these definitions, Eq. (17) may be written as
T -+ 2 T T
K 1 K55 exp[ 00a 00 + xTa-c 0 o0 0 - 0-X c-x 2 Qx]
QxI . (21)
13
·I··^____ I_III1IYIIII______I_ I I .I _-
Substituting Eq. (21) into Eq. (15), and ignoring the factor K1K 5 exp [--c02] since it will appear
in both the numerator and denominator of the likelihood ratio and thus be canceled out, we find
1 T -1 2T -1 T -1
= exp [ 0 a0 + aT-aQ a+ TQ c Q-tc TQ a] (22)
Thus, it follows from Eq. (22) that the likelihood ratio Eq. (2) becomes
H1
a0 cTQ-a >
a- Q a < 0 (24)
Ho0
g_2
Q-1 A
-Q c (25)
g2
H1
14
A 90
9 (27)
91
MATCHED
FILTER
Ht
<o
v
Ho
At this point, we would like to determine those tap-gain settings which optimize the equalizer
performance, minimizing its probability of error. For the special case where the sum of the
3
sidelobe magnitudes is less than unity, Aaron and Tufts have done this by first minimizing the
output noise variance for a given set of sidelobes, and then working through a specific example,
employing a search in tap-gain space to arrive at those settings which minimize the probability
of error. This approach was feasible only because they consider small numbers of sidelobes
and taps, thus requiring summation over only a small number of possible message sequences.
In realistic applications, as we discuss further in Sec. IV, the number of terms involved in such
a probability-of-error computation becomes prohibitively large, and we therefore must be con-
tent with some other method, and with less-than-optimal tap-gain settings.
Instead of attempting to minimize the probability of error directly, authors usually attempt
to minimize the intersymbol interference appearing at the receiver output. Two distortion
measures are currently in use. With the mainlobe assumed unity, and the summations over the
sidelobes, these measures are
D = E
0L qi (28)
15
where the qi are the output sidelobes resulting from a single pulse transmission, as shown in
9'
Fig. 8. DiToro uses the D a criterion in his ADAPTICOM receiver, while Lucky uses the
D: criterion, although he asserts there is probably little difference in terms of performance
using either criterion. In this report, we adopt a modified Da criterion under which we include
the additive noise appearing at the output, and find those tap-gain settings which minimize the
total output distortion.
qn 13-42- 107461
-M
-M
iT 1
2I
q-2
T
qI
q2-
1
T t
M
qM
a single to = +1 pulse transmission.
We first determine an expression for the output distortion in terms of the TDL gains. For
a single 0 = 1 transmitted pulse, synchronous sampling of the matched filter output renders a
sequence of samples
and
The TDL output is then a sequence of samples given by the convolution of the zk sequence with
the tap-gain vector g:
Z =
k-mgm £ k-mg m + E Wk-mg m
m m m
qk + k (33)
where we have defined the first sum on the right-hand side to be qk, and the last sum to be nk,
as indicated. In transmitting an infinite sequence of pulses modulated by the k (which we have
previously assumed to be independent, of unit variance, and zero mean), the output distortion
at any sample time, say, the zeroth sample time, will be given by:
2 2 2
Output distortion = q-q 0 + no . (34)
k
16
I _
Normalizing the tap gains such that q = 1, then to minimize the average output distortion, we
must minimize the quantity
+ n 2(35) 2
qk + no
k
q2=
k
q- , (=(Z ·Pk-mgm)
,kmm (Z (z k-vgv)
(k-vgv)
k k m v
and
N
NO gg ~-
y(37)
gmPm_vgv
m v
the latter step holding since, from Eqs. (31) and (32), it follows that
2 NOPmv (38)
and
Y mv =P n-v
m-v (40)
then substituting these into Eqs. (36) and (37), using the tap-gain vector notation g defined in
Eq. (27), we may write Eq. (35) as
Under the constraint that the mainlobe is unity, which we may write as
T T
g S° = _ = 1 (42)
where we have defined the column vector
17
--X--I·IP-
1 _·II~~~~~~--IP1-I~~~~.IIIYY
__.~~~_1111111_11 ------
1
(P0 (43)
'P
L.
then our problem is to minimize Eq. (40) subject to the constraint Eq. (41). Introducing the
Lagrange multiplier , we may do this by minimizing the quantity
in order to find the tap-gain vector g which is optimal under our distortion criterion. It is
straightforward to show that
A 1
=[X + 1 NOY] g + Op (45)
aJ
ag
L
and, further, that aJ/ag is zero for
1
[X + NY] (
(46)
g= o [X NOY] _
18
X+ 1
2
N 0-Y (47)
is positive definite, and its inverse exists (Ref. 14, p. 46). Since the inverse of a matrix is itself
unique, it follows that the solution Eq. (46) exists and is unique. Moreover, the facts that
2
= 2[X + NY (48)
2 -- 2 oY-
and that the matrix of (47) is positive definite, are sufficient conditions to guarantee that our
solution Eq. (46) does indeed render a minimum of Eq. (44) (see Ref. 15, p. 227).
Having thus arrived at the tap-gain settings which minimize the total distortion at the output,
we next want to determine their asymptotic behavior as the length of the TDL becomes arbitrarily
large. We define a unity vector u:
1 (49)
0
Noting that
(YU) k = km m =
kO (50)
m
and
(51)
(YY)mv= Z YmkYkv = '°m-k ~° k-v = Xmv
k k
Yu= p Y 1= = u (52)
and
YY= X (53)
Using Eqs. (52) and (53) in Eq. (46), and ignoring for the moment the denominator which is merely
the normalization, we find asymptotically the solution
19
We next want to study this asymptotic result still further. Rewriting Eq. (54) in terms of a sum-
mation, we have
+ 6 6
[Pi-j 2 N ij] gj= i0 (55)
J
Subtracting the j = 0 term from both sides, and normalizing the tap-gain vector such that gO = 1,
we find the tap gains must satisfy
1
+
['i-j 2 No06ij gj = 'i (56)
j#0
for all i # 0. Dividing through by No, and noting by comparison of Eqs. (8) and (31) that i/No =
bi, then Eq. (56) becomes
[Y + 2 NoI] g = (58)
From Eq. (40) we note that Y is a Toeplitz matrix, since the elements along any of its diagonals
=
are the same, Yij Yi+ ,j+t (see, for example, Ref. 16). Thus, asymptotically, Eq. (58) be-
comes equivalent to convolving the row vector
with the tap-gain vector g to obtain the vector u. Therefore, this may be written as a set of
equations
e(w)
r<P_ eJ-ljCMW (62)
m
and
20
it follows that we may solve Eq. (60) by first setting
G(co) =1 (64)
4(w) + - N0
The relation Eq. (64) was obtained previously by George,1 7 except for an arbitrary gain factor
he incurs by neglecting normalization. George arrived at this relation somewhat more directly
by starting out initially with an unconstrained linear receiver. Later, Coll t 8
adopted the re-
sulting conventional receiver structure and, using a variational approach as George had done,
determined the tap-gain equations for minimizing the output distortion for TDL of finite length.
Our formulation led directly to the solution for TDL of finite length, and then asymptotically to
George's result. The approach we present here enabled us to establish the existence and unique-
ness of the solution, and to relate it asymptotically with the tap gains obtained in our original
derivation of the conventional equalizer structure.
As explained in the Introduction of Sec. I, we will study the performance of the sub-optimal
conventional equalizer in detail in Sec. IV, but first, in Sec. III we will consider another sub-
optimal receiver - the decision-feedback equalizer.
21
1-1-1l1--11 - -111__--_1
- --- I-I-II
III. EQUALIZATION USING DECISION FEEDBACK
We saw in Sec. I that previous applications of decision feedback in equalization have been
concerned, indirectly at least, with effectively measuring the equivalent channel in order to
achieve tap-gain settings for the conventional equalizer we derived in Sec. II or its cascade
equivalent. Next we derive a new equalizer structure in which the decision feedback plays a
different role. We then optimize the parameters of this new structure and, finally, we briefly
compare its sidelobe suppression properties with those of the conventional equalizer.
Our problem is the same as that stated in Sec. II, namely, to decide between the hypotheses
Hi and H 0 that 40 is +1 or -1, respectively, given a receiver input
we found the optimal equalizer structure of Fig. 6. We now make two additional assumptions
which will lead us to a sub-optimal receiver structure:
(4) k are Gaussian random variables of zero mean and unit variance
for k > 0, and
(5) Ok are known for k < 0 (i.e., error-free decision feedback).
As we observed in making a similar assumption in Sec. II, Assumption (4) renders our model
inaccurate for binary AM or PSK systems in which k = +1 or -1 for all k, while Assumption (5)
is valid for decision-feedback equalizers only in the absence of decision errors.
Under our Assumption (5) that , as defined earlier by Eq. (3), is known correctly via
decision feedback, the optimal receiver computes the likelihood ratio
p[r(t)l -, 0 =1]
A =
p[r(t)l-, 0 = -1
- ' ]
Sp[r(t)l 0 t p(+) d + =1
(67)
f p[r(t)l 0, , ] p( ) dl _ (67)
23
---
---- ···------------ ·----------
"I'-""`-
exp [Oa0 + ma m k2 k-m
kmb
m>0 k<0 m>0
-2~o Z kbk
-2tOZ mbm -
3 ijmb j-m (69)
k<0 m>0 j>0 m>O
Now, under Assumption (4), we may write
p(+) = K 2 exp =
-2
[ m K2 exp - 2 jmm0
.j 7
m>0 j>0 m>0
Thus, by combining Eqs. (69) and (70), the numerator of Eq. (67) may be written as
24
exp 0 a0 - o Z 4kbk exp- Z Z j(bjm + 2 6
jm) m
k<O j>0 m>0
+2 m( - bkm- --
b)j d) (71)
m>O k<O k0=1
By completing the square in the exponent of the integrand of Eq. (71), it is straightforward to
show that this numerator of A is
exp [ 0a 0 -2 0 b +
kkbkZ (2 aj- bj kk- b) Pj
k<0 j>0 m>0 k<0
(m Z b m-4- 0 m | (72)
n<O 0
where the Pjm are the elements of the matrix P defined by
1
PA R- (73)
9 Z P jm bm (76)
g jm
m>O
24
fkA 2bk + 2 gjbj k (77)
j>o
Hi
As shown in Sec. II [see Fig. 4(b)], the a are sufficient statistics which may be generated by
matched filtering to the equivalent channel, sampling the output synchronously at the baud rate,
and passing the samples into a TDL. Note, however, that with the present sub-optimal equal-
izer, only the a for j > O0enter into the decision rule, just as we noted is true of the optimal
equalizer whenever the previously received baud modulations k are known. Whereas the effects
of these earlier-baud ~k were accounted for in the optimal equalizer structure by the C(s) con-
stants of Eq. (13), which themselves undergo nonlinear operations before effecting the output,
they enter in this sub-optimal equalizer as a weighted sum, the second term of Eq. (78). Thus,
the overall decision rule of this sub-optimal receiver may be implemented as shown in Fig. 9.
SAMPLEAT 13-47-047-
INPUT
25
I I-·-III _I -- -
q = signal component of the forward-TDL output at the mth sample
time, when a single ~0 = +1 baud is transmitted (79)
=
A typical response of the decision-feedback equalizer to a single transmitted baud of 0 +1
is shown in Fig. 10. It is always an asymmetrical waveform, having M more samples occurring
before the main sample (which is denoted sample
qo 113-42-s07481number 0) than after it, where the MF output has
q' ' 2M + 1 nonzero samples. This is in contrast with
the typical output from the conventional MF-TDL
equalizer, which was seen in Fig. 8 to always ex-
9" -2 ? Thibit symmetry about the main sample.
Before we can proceed to determine the optimum
choices of g and f under our minimum-output-
q3
sample-variance criterion, we must first under-
Fig. 10. Typical output sidelobes from decision- stand the effect of the decision feedback on the
feedback equalizer for a single 0 = +1 pulse output distortion. Consider the signal component
transmission. out of the forward-TDL at the first sample time:
The contribution to this component, which is due to the bauds for which decisions have already
been made (that is, on all the k up to and including 0 ) , is then
Next, consider the output of the feedback-TDL at this same first sample time. With the use
of Eq. (77) for the fk' this becomes
26
and, if we let k* = k + 1,
£ kk+1 2
£ fbk,,- + gjibj+lk. k*
k<O k*<O j>O
Here, we have used bk*,1 = b i k* and our earlier definition, go A 1. Thus, we see from
comparing Eqs. (87) and (89) that in the absence of decision errors the feedback-TDL output is
exactly the same as the contribution to the forward-TDL output at the first sample time, attrib-
utable to past bauds, hence, there is no net contribution to the distortion from those bauds upon
which decisions have already been made.
In abandoning Eqs. (76) and (77) in order to optimize the tap gains under a minimum-output-
distortion criterion, we wish to retain the nature and purpose of the feedback-TDL as seen above
with the optimal solution - namely, to cancel out sidelobes attributable to past bauds. In the
absence of decision errors, then, with the feedback-TDL tap-gain vector f suitably chosen, it
is clear that the output distortion we want to minimize is given by
Under the constraint that the main sample be unity, we may include q0 in the summation to find
that
=
, qm= , , gjXjkgk g (91)
m(0 j) kO
2 gjYjkgk = Y
Yg (92)
j) kO
Thus, under the constraint of Eq. (42) that the mainlobe be unity, we proceed as in Sec. II to
minimize the quantity
J=gT[x+ YJ g + x( gT ) (93)
over the choice of the forward-TDL gain vector g. The unique solution is again given by
X + NO Y _p
g (94)
TIX+ - 2Y _
This result appears formally identical with that obtained for the conventional equalizer, Eq. (46).
The differences are in that we have redefined the vector qo and the matrices X and Y in Defi-
nitions (81), (82), and (83), and now their subscript ranges differ from the ranges used in study-
ing the conventional equalizer. Thus, while (p was symmetrical and X Toeplitz for the conven-
tional equalizer, these properties do not hold for the _o and X for the decision-feedback
27
_____ _~~~ II L ·
equalizer. Both of these differences result from the fact that sidelobes occurring after the
mainlobe are ignored in the decision-feedback equalizer. In Definition (82), for example, this
is reflected in that the summation is restricted to m < 0, whereas the corresponding summation
Eq. (39) of the conventional equalizer has no such restriction.
The proper choice of the f vector now follows directly from Eq. (77), except that we have
dropped the 2/N 0 factor common to the terms of Eq. (78) (that is, the aj contained this factor,
while it has not been included in the MF output in the present discussion), and thus the 2b k be-
come replaced by the cok of Definition (80):
=
Yjk j-k for
jk < 0 (96)
[note that it is the range of k which distinguishes this matrix from the Y matrix of Definition
(81)], then the feedback-TDL tap-gain vector may be conveniently written as
f= g (97)
Thus, once the sampled channel autocorrelation function and the additive noise level have been
specified, we can use Eqs. (94) and (97) to determine the tap gains of the minimum-output-
distortion decision-feedback equalizer. Moreover, using the same arguments we presented for
the conventional equalizer, we can show that the solutions to these equations exist, are unique,
and render a minimum of Eq. (90).
We want next to compare the operation of the decision-feedback equalizer of Fig. 9 with that
of the conventional equalizer structure shown in Fig. 7. Since in actual implementation, equal-
izer cost is directly related to the number of delay elements and amplifiers required, then our
comparisons will be based on equalizers having the same number of taps. The conventional
equalizer, of course, uses its allotment of taps on its single TDL, while the decision-feedback
equalizer must devote part of its taps to the feedback-TDL and the remainder to the forward-TDL.
Before proceeding, we want to consider the number of taps required on the feedback-TDL.
We may rewrite Eq. (95) as
=
f-k Pmgm-k for k > 0 (98)
m>-k
and thus, for a channel where the maximal dispersion is over, say, N bauds (that is, where
=
0on
0 for n > N), then from Definition (83) it is clear that
Thus, at most, N - 1 taps are required in the feedback-TDL of the decision-feedback equalizer,
and any additional taps which are available may be employed in the forward-TDL.
The sampled MF output for a single pulse transmitted is always symmetrical, thus it follows
from Eq. (38) that a conventional equalizer having an odd number of taps always has a symmetri-
cal tap-gain vector. Hence, a (2M + )-tap conventional equalizer really has only M degrees of
28
Ad L5-42-1o?4.] freedom, assuming that the centertap gain is
0~~~~~~f-zils
normalized to unity. With these M degrees of
freedom, the conventional equalizer must si-
multaneously attempt to suppress the sidelobes
attributable to both past and future baud trans-
missions. On the other hand, as' we noted above,
the decision-feedback equalizer feedback-TDL
requires only N - 1 taps for a channel with a
Fig. 11. Example sampled channel autocorrelation
function for illustrating sidelobe suppression in con- maximal dispersion of N bauds, leaving 2
ventional and decision-feedback equalizers. 2 - N taps and degrees of freedom available for
the forward-TDL. Moreover, the forward-TDL
gains are computed without regard to the qm for m > 0, and only the qm for m < 0 are taken into
account (see Fig. 10). Thus, the forward-TDL need only be concerned with suppressing the
sidelobes attributable to future bauds, that is, those bauds upon which decisions have not yet
been made.
To illustrate the advantages resulting from the greater degrees of freedom available to the
decision-feedback equalizer, we will consider the sampled channel autocorrelation function
shown in Fig. 11. Using a 21-tap TDL in the conventional equalizer, with the tap gains given
by Eq. (46), the tap-gain vector has a limiting solution with increasing SNR and, accordingly,
there is a limiting sidelobe behavior at the output
13-2-4o1sol
for a single pulse transmission, as shown in
Fig. 12(a). The forward-TDL output of the equiv-
alent decision-feedback equalizer, with its tap
gains computed according to Eq. (94), is shown
in Fig. 12(b). In this example, the conventional
equalizer has M = 10 degrees of freedom com-
pared with 2M + 2 - N = 17 degrees of freedom
available to the forward-TDL of the decision-
feedback equalizer. The resulting sidelobes qm I II I I I
for m < 0 for the decision-feedback equalizer are
seen to be considerably smaller under either a CONVENTIONAL
EQUALIZER
(a) CONVENTIONAL EQUALIZER
Da or D distortion criterion [defined in Eqs. (28)
and (29), respectively] than are the sidelobes of
the conventional equalizer. As we discuss fur-
ther in Secs. IV, V, and VI, this greater sidelobe-
suppression behavior results in a large improve-
ment in the performance of the decision-feedback
equalizer over that of the conventional equalizer.
As expected, the sidelobes qm for m > 0 at the
output of the decision-feedback equalizer forward-
TDL are very large, since they were not involved
in the tap-gain optimization. In the absence of (b) DECISION-FEEDBACK
EQUALIZER
decision errors, these large sidelobes are ex-
Fig. 12(a-b). Limiting sidelobes with increasing
actly subtracted out by the feedback-TDL com- SNR of 2 1-tap conventional and decision-feedback
ponent of the output, as we observed previously. equalizers, applied to channel of Fig. 11.
29
30
IV. PERFORMANCE OF THE CONVENTIONAL EQUALIZER
We now evaluate the performance of the conventional equalizer structure derived in Sec. II,
with a view to comparing it with the performance of the decision-feedback equalizer derived in
Sec. III. Such a comparison, however, necessarily involves consideration of a specific channel,
or class of channels. We describe next the class of channels which we have chosen (the maxi-
mal distortion channels), after which we present an algorithm for the efficient evaluation of the
conventional equalizer performance; finally, we present the results obtained by applying this
new algorithm to the maximal distortion channels.
N-1 N-1
- jn o
D(@) A Z dne = 1+ 2 dn cos(nwo) (100)
n= -N+ 1 n= 1
be non-negative over the interval w E (0, r). (The transform must be non-negative for all o, of
course, but D(w) is symmetrical about the origin and of period 2r, so we only need consider the
interval indicated. )
-3
At this point, we want to visualize an (N - )-dimensional " sidelobe space," with the dn as
the coordinates. We then define that portion of this sidelobe space within which D(w) is non-
negative over (0, 7r) as the "region of realizability." This region always includes the origin, since
all the sidelobes may be set equal to zero, corresponding to a nondispersive channel. Moreover,
since it is a well-known property of autocorrelation functions that they attain their maximum at
the point corresponding to the center sample d in Fig. 13, which we have normalized to unity,
it then follows that the region of realizability is bounded, being within a hypercube having sides
of length two centered at the origin. By its very definition, it follows that, at points outside the
region of realizability, D(w) becomes negative. Since, from Eq. (100) we observe that D(w) is a
continuous function of the dn , then it clearly must just vanish at its minima on the boundary of
31
the region of realizability. Moreover, the quantity we want to maximize subject to realizability,
D:, is recognized to be constant over hyperplanes in each hyperquadrant of sidelobe space. Since
D: increases with the distance of these hyperplanes from the origin, it is clear that any point
maximizing Do must lie on the boundary of the region of realizability, for the maximum DB will
be attained on the hyperplane which is just tangent to the region of realizability. From this and
our above result, we conclude that D(w) must vanish at its minima for the desired maximal dis-
tortion channel.
We next consider the minima of D(w). Let wi be a frequency at which D(w) has a minimum.
Then, wi must satisfy the two equations
N-i
)
D'(wi =-Z ndn sin(ni)= (101)
n=1
N-1
2
D(wi) =-2 n dn cos(nwi ) >0 ( 02)
n=1
where Eq. (101) requires that coi correspond to a critical point, and Eq. (102) insures that it is a
minimum rather than a maximum or inflection point. Now, since
where p(n) [x] denotes an nth-order polynomial in x, then clearly two of the critical points are
= =
always at wi 0 and woi 7r. Now, n N - 1 in Eq. (101); thus, from Eq. (103) the highest order
polynomial in cos (i) encountered there is of order N - 2. Moreover, since cos (i) is single
valued over the (0, 7r) interval of interest, then this polynomial will render exactly N - 2 addi-
tional critical frequencies wi, giving a total number of N critical frequencies for the Nth-order
channel of Fig. 13. N/2 of these, or the largest integer in N/2, correspond to minima of D(w).
Now, imposing the constraint arrived at earlier that D(w) vanish at its minima for the max-
imal distortion channel, we want to maximize the quantity
[N/2]
J = D + iD(wi) (104)
i=l
where we have introduced the Lagrange multipliers Xi , [N/21 denotes the largest integer in N/2,
and the summation is seen to be taken over all those o.1 corresponding to the minima of D(w).
Noting from Eq. (29) that
aD
aD
9 (105)
3d = 2 sgn(dn)
n
D(.i) a3W.(
ad 1 = 2 cos (n i) + ) (106)
n w= Wi n
where, at the minima points wi, the quantity aD()/8w vanishes, we may combine Eqs. (104) to
(106) to find aJ/adn. Doing this, and recalling our constraints, we find the set of equations
which the maximal distortion solutions must satisfy:
32
= sgn(d + Xi cos (nwi)= (107)
n i=t
N-i
D(wi) = + 2 dn cos(nwi)= O (108)
n=1
Next, we observe that if d,.. . dN i is a solution to Eqs. (107) and (108), then, from the
relation
N-1
D()A- +2 dncos(now)
n=1
N-I
= + 2 , (-)n dn cos [n()r-l-A D(wr- w) (109)
n=1
d = (-1)n
n dn (110)
also renders a maximal distortion solution, since from Eq. (109) it is seen that D(w) is simply
D(w) reversed within the interval (0, 7r). Thus, we conclude that the odd-numbered sidelobes of
a maximal distortion solution may all be reversed in sign to obtain a second maximal distortion
solution. We henceforth refer to this as the "symmetry" property with respect to the odd-
numbered sidelobes.
We now want to apply all of the above, and to determine the class of maximal distortion
channels. Considering first the simplest case where the channel is spread over two bauds
(N = 2), we have a single sidelobe d1 . Because Eq. (107) involves sgn(d 1 ), there are two situa-
tions to consider. However, one of these can be obtained from the symmetry property shown
above, thus we need only consider the case where d1 is positive. In this case, Eqs. (107), (108),
(101), and (102) become, respectively,
aJ
=
a =1 + cos ( 1) 0
X1=1
cos(wl) = -1
sin(ow) = 0
d1t I2 (112)
33
Thus, by using the symmetry property, the two maximal solutions are d 1 = ±1/2, both rendering
a distortion D = i. The corresponding sampled channel autocorrelation functions and a cycle
of their transforms are shown in Figs. 14(a) and (b), respectively. These will be discussed
further below.
13 -2-10152
* *
I -L-CC-
2
-1i 11
*- *v
- -
i -- *
- v 0 v
2
(a) (b)
For the N = 3 case, there are four situations, two of which may be found from the symmetry
property. Remaining then are case (1) where di and d 2 are both positive, and case (2) where d
is positive and d 2 is negative. For case (1), the equations we must solve become
aJ
)
ad i=
1 + cos(o = 0
ad i + X cos(w) = 0
= (11 3)
D () -2d 1 cos (w)- 8d 2 cos (2wl) > 0
A =2
cos
) = 2
Cos (w ) = Cos (2 ) 2
) =
sin ( -sin (w 1 ) = 2
21
d1 = 3
1
d = (114)
2 3
34
corresponding to a D: = 2. For case (2), all the above equations still hold except that for
aJ/ad 2 , which becomes
dz -- + cos(2 1 ) = 0 (115)
k= 1
=
cos () -cos (Zw1 ) = -1
sin(wl)= sin(2w) = 0
2d 1 -2d 2 =1 (116)
But this last equation simply states that D 1, which is less than that obtained with case (1).
Applying the symmetry property, it therefore follows that the maximal distortion solutions are
d = + 2/3, d 2 = 1/3. The corresponding sampled channel autocorrelation functions and a cycle
of their transforms are shown in Figs. 15(a) and (b), respectively.
From Figs. 14(a) and 1 5(a), one might suspect that perhaps one of the maximal distortion
solutions always has a triangular envelope, as indicated in Figs. 16 (a) and (b) for the N = 2 and
N = 3 cases considered above. This may be investigated for arbitrary N by considering the
triangular channel autocorrelation function SOhh(T) shown in Fig. 16(c). Its conventional Fourier
transform is given by
__]
142-_oT5 1
//
'%%%% (a)
/
-2 -1 0 1 2
./'T./
I I| |(b)
,
-3 -2 - 0o 1 2 3
(T )
hh
-N N (
Fig. 15. Third-order maximal distortion channels: Fig. 16. Autocorrelation functions with triangular en-
(a) sampled autocorrelation functions, and (b) cor- velopes: (a) second-order maximal distortion channel;
responding frequency spectra. (b) third-order maximal distortion channel; (c) contin-
uous Nth-order channel autocorrelation function.
35
hh(W) = 'Phh() ejWT dT = N [i (N/2)]2 (117)
as sketched in Fig. 17(a). The equivalent to the sampled version of Fig. 16(b) may be obtained
by multiplying it with an impulse train whose Fourier transform, also an impulse train, is
shown in Fig. 17(b). The transform of the resulting sampled channel autocorrelation function
(a)
-4r --2T 0 2r 4r
(b)
may then be found through convolution of the waveforms in Figs. 17(a) and (b). The resulting
transform D(w) has, by inspection, the following properties:
It therefore follows that sampled channel autocorrelation functions with triangular envelopes
satisfy conditions of Eqs. (101), (102), and (108). To show they are indeed maximal distortion
solutions, we must show that Eq. (107) is satisfied as well. As explained earlier, only the min-
ima occurring within the interval (0, 7r) need be considered, which from property (1) above means
summing over 1 i . N/2 in Eq. (107):
[N/2]
ad - + X.icos (27rni/N) = 0 (118)
n i=
for 1 < n < N - , where we have applied sgn(dn) = 1 for the present case. Moreover, since
then, to prove that the sampled channel autocorrelation function having a triangular envelope
vanishing at = N is the maximal distortion channel, we need only to show that a set of Lagrange
multipliers Xi exists to satisfy Eq. (ii8) for n = . .... [N/Z].
36
If we define a matrix A with elements
-1
where we recall that [N/2] is defined to be the largest integer in N/2. A solution to Eq. (120)
exists whenever A is nonsingular, having a nonzero determinant. While we have not been able
to establish this for general N, it appears true, as shown below for N = 2 through N = 7:
Granting that sampled channel autocorrelation functions having triangular envelopes (such
as we have drawn in Fig. 18) render maximal distortion (as we showed above, at least for chan-
nels with dispersions up to N = 7 bauds duration), we find from the symmetry property that a
second maximal distortion solution may be obtained by changing the signs of the odd-numbered
sidelobes. From Eq. (46), however, it is easily established for the conventional equalizer that
its optimal tap-gain vectors corresponding to these two choices of maximal distortion channels
O= 13-42-OTS,1
Fig. 18.
-
. - -
--N
-
-N+1
,TI I I ITI,
. .
...
. . ....
-3 -2 -1
I L
0 1 2
3
.
...
.
N-1
____
S **
Z r
37
are identical, except that the polarity of their odd-numbered gains is reversed. Moreover, this
renders output sidelobes which are identical except for a polarity reversal of the odd-numbered
sidelobes. Consequently, the output distortion with either choice of sampled channel autocor-
relation function is the same. From Eq. (94), we see that this is also true for the decision-
feedback equalizer; thus, in the cases to be considered later in this Section and again in Sec. V,
we will arbitrarily choose the first case, where the dn are taken as positive.
We next want to determine the performance of the conventional equalizer derived in Sec. II.
=
As we noted previously, the input to the conventional equalizer TDL for a single 0 1 trans-
mitted pulse is the symmetrical sampled channel autocorrelation function, resulting in a sym-
metrical tap-gain vector and symmetrical output sidelobes, as sketched in Fig. 8. If there are
a total of 2M such sidelobes (M on either side of the mainlobe), then in making a decision on,
say, 0 when a continuous message sequence has been transmitted, these sidelobes are weighted
by the k to render the intersymbol interference distortion D appearing at the output:
where the prime on the summation denotes that the k = 0 term is to be omitted from the sum.
For any given distortion D and output additive noise variance , we may determine the con-
ditional probability of error from
where
erfc(z) = e2 dt (123)
is the complementary error function. Thus, noting that, for a given set of sidelobes, D depends
4
only upon the k' it is clear that we could (at least conceptually) determine the probability of
error via the relation
+ ) + )
p(error)= , , p(errorl -, p( -, (124)
- +
where we have here modified Definitions (3) and (4) to cover only the desired M bauds into the
past and future:
f - {~kli k M) (i26)
and where
38
The difficulty with this straightforward approach (adopted for a simple case in Ref. 13) is
that the summation of Eq. (124) involves a total of 2ZM terms which, for practical channel dis-
persions and equalizer lengths of interest, can become prohibitively large. For example, in
one of the cases considered below, M = 19, corresponding to 2 38 = 2.7 x 101i terms to be in-
cluded in the summation of Eq. (124).
We may simplify matters somewhat by first rewriting Eq. (121):
M
+ )
D Z (q-kk .qk-k (128)
k=1
Then, since qk = qk from sidelobe symmetry, it follows that the kt h term of Eq. (128) may as-
sume three values:
1
2qk with probability
1
(q-kk + qk -k) =I with probability 2
bk Zlqk| (130)
0 with probability 1
77k =
M
D = bklk (132)
k=i
where we have defined ~ = {k } The summation of Eq. (133) has 3M terms, since the M dif-
7 M
ferent k may each assume the three values indicated in Eq. (131). Since 3M < 2 2M = 4 , it is
seen that Eq. (133) represents a large reduction in computation over Eq. (124), indeed by a fac-
tor of (4 / 3 )M = 235 for our M = 19 case. Unfortunately, however, 39 = 1.1 x 109 is still far
too large a number of terms to sum in practice. Therefore, we next derive a new algorithm
which, as we will see later in this Section, reduces further the summation requirements to the
point where obtaining arbitrarily close upper and lower bounds on Eq. (133) becomes computa-
tionally feasible.
39
We begin by rearranging the sidelobes of Eq. (130), defining a new set of sidelobes denoted
by bi, b2 , ... bM which correspond one-to-one with the bk of Eq. (130), but where now they are
ordered such that
These sidelobes, just as those of Eq. (130), are all positive, corresponding to the doubled mag-
nitudes of the original set of sidelobes qk except for ordering. Also, defining = {I' 72, ....
to be the corresponding rearrangement of , we may write Eqs. (132) and (133) as
D= bk77k (135)
k=1
Because the weightings bk in Eq. (135) are non-increasing and usually decreasing with k, we
would expect that many of the r7 appearing in Eq. (136) will have distortions D essentially de-
termined byotheir first few components. If this were strictly true, we could compute the com-
bined contribution to Eq. (136) from all 77 having identical first few components, thereby reduc-
ing the overall number of terms to be included in Eq. (136). Since the latter components of 1
cannot be legitimately ignored altogether, however, we instead consider upper and lower bounds
on their contribution.
7
Leht . dK+1. M where the k have been specified and the Ok
are arbitrary. We may then define quantities
K M
DL(K)-a bk k + E bk (137)
k=1 k=K+1
K M
Du(K)A Z bkT - Z bk (138)
k=1 k=K+1
which correspond to the largest and smallest distortions which any such 77* can cause. The
subscripts L and U are adopted because use of these distortions in Eq. (122) renders "lower"
and "upper" bounds on the conditional probability of error for any such :, respectively:
Next, we let 77 E 71' denote that the first K components of 77 are identical with those of 77 .
Then, from the discussion above, we may write for any such r7:
40
_I
K M
Z p )S p(H 7 Pp(nj)
7
77ET k=1 r lK+1 iM j=K+1
K
lH (Ij) (142)
k=1
then, if we define
K
W(K) = H p( ) (143)
k=1
to be the probability of any such Tqtaking on these initial K components, from Eqs. (141) to (143)
we may write
The inequalities of Eq. (144) provide lower and upper bounds on the total contribution to
Eq. (136) from all those 7r E 7*. If these bounds are sufficiently close in some sense (which we
consider further below), then we need not bother summing over all the additional components of
these _~ (that is, over the 77k for k > K), for the result of such additional computations cannot
change the net contribution to Eq. (136) from all such r by more than the bounds of Eq. (144)
allow.
Since we may rewrite Eq. (136) as
where the first summation is over distinct rT:-', then from Eqs. (144) and (145) the desired prob-
ability of error for the conventional equalizer may be bounded above and below via
where in general, of course, the value of K required to bring the bounds of Eq. (144) sufficiently
close together will differ for the different choices of 77*. The actual computation of the bounds
in Eq. (146) is accomplished through the sequential searching procedure described below.
Consider the tree-like structure shown in Fig. 19. Starting at the left, we trace through the
tree to the right, determining a particular sequence 77i, . . . by our choice of paths. At each
node, we want to stop and compute the lower and upper bounds of Eq. (144). To accomplish this,
we need to keep track of W(K), DL(K), and DU(K) as we progress through the tree. This is done
conveniently by means of the recursive relationships
41
/ 13- 2-0571 where K is the number of branches separating the
present node from the first node, and where we
define the initial values
W(O) 1
M
D(0) 1 bk-Du(0) (150)
k=t
42
I
101
y 8'
Z
4 W
X
t
- Y
J
4
!i
0
-
I
Ir al
W
U) 6
U Z
4
C.)
z 4
w
4 t
o CD
i
0
oa. 4'
0
X
0 U.
V}
V)
O
W
.-
W
z
2 Q
a z
U
(
0.
-16 -4 8 20 32 44 56 -16 -4 a 20 32 44 56
(a) (b)
Fig. 20. Efficiency of error tree algorithm in determining performance of conventional equalizer
applied to maximal distortion channels: (a) 1 -tap equalizer, (b) 21-tap equalizer.
close together are small, resulting in a brief and efficient search of the tree. For intermediate
SNR this no longer holds, for there are many combinations of message sequences and additive
noises which can lead to an error, thus requiring that the algorithm search out a much larger
percentage of the tree. Finally, at high SNR, unless the intersymbol interference is rather
strongly opposed to a correct decision, an error is unlikely to occur and, since there are a
relatively small number of sidelobe combinations leading to strong opposition, then, here again
(as at low SNR), the algorithm need only search a small percentage of the total number of pos-
sible paths in the tree.
We observe at this point that a possible alternative to our error tree algorithm approach
would be to determine the performance of the conventional equalizer through digital simulations,
as we do in Sec. V when considering the decision-feedback equalizer performance. This would
work out well at low and intermediate SNR, but at high SNR, when the error rates are generally
low, this alternative approach would require simulations of prohibitively long duration. Thus,
it is in the high SNR cases where the algorithm proves most useful, for it efficiently and accu-
rately renders the bounds required in Eq. (146).
Section C below presents the results obtained through applying our algorithm to conventional
equalizer performance calculations for the maximal distortion channels described in Sec. A.
43
Sec. A above. In the following discussions, we will denote by L the number of taps used on the
conventional equalizer TDL, frequently referring to L as the TDL length. We denote by N the
order of the maximal distortion channel, where N is the number of bauds over which the chan-
nel dispersion extends. Also, we may occasionally refer to the situation where a maximal dis-
tortion channel of order N is equalized using a TDL having L taps as the "N-L case." We con-
sidered several different cases, for various N, L, and SNR, as described below.
For each case studied, we first performed the calculations indicated by Eq. (46) to arrive
at the optimal tap-gain settings of the conventional equalizer under our distortion criterion.
That is, we computed those gains which minimized the total distortion encountered at the output,
the sum of the sidelobe energy and output noise variance. With these tap gains, we then calcu-
lated the additive noise variance appearing at the output. Next, we determined the output side-
lobes for a single pulse transmission, and the corresponding distortion measures D and D.
Finally, using this set of sidelobes and the output noise variance, we applied the error tree al-
gorithm developed in Sec. B to determine the performance of the conventional equalizer.
In this Section, we will be chiefly concerned with the effects of N, L, and SNR upon the
conventional equalizer performance, its probability of error. The remaining data we obtained
on tap-gain and sidelobe behavior will prove of interest as well in Sec. VI, when we compare the
conventional equalizer operation with that of the decision-feedback equalizer.
The results of our performance calculations are summarized in Tables I to IV, which con-
tain the probability-of-error data obtained with the maximal distortion channels of orders N = 2
to N = 5, respectively. As these tables show, we permitted the SNR to range between-16 and
+50 db for each combination of N and L chosen. This rather broad range of SNR enabled us to
observe the performance of the conventional equalizer at two extremes: (1) at low SNR, where
the performance is essentially noise limited and independent of the intersymbol interference,
and (2) at high SNR, where it becomes noise independent and limited only by the TDL length, as
we discuss further below.
The data of Tables I to IV are presented in two sets of curves. The first set, comprising
Figs. 21(a) to (d), shows performance vs SNR for the different TDL lengths L, with a separate
figure for each order maximal distortion channel. We have also included in each of these fig-
ures the curve corresponding to the nondispersive N = 1 case. This latter curve, of course,
provides a performance bound, representing the best performance we could obtain for any order
maximal distortion channel, if we were to separate the transmitted pulses sufficiently in time
that the channel dispersion no longer caused intersymbol interference. The second set of curves,
consisting of Figs. 22(a) to (d), plots performance vs L for different values of SNR, again with
a separate figure for each order maximal distortion channel. Other combinations are, of course,
possible (e.g., plotting L vs SNR for different performance levels and N), but these appear to
lend little additional insight into the conventional equalizer operation over that already afforded
by Figs. 21(a) to 22(d).
In Fig. 21(a) for the maximal distortion channel of order N = 2, we note that at very low
SNR the performance depends upon the SNR only, and becomes independent of the length of the
TDL employed. This, of course, is true because at very low SNR all the tap gains approach
zero except that of the centertap on the TDL; thus, whether or not additional taps are available
is unimportant, since their gains would be virtually zero anyhow. These same comments are
seen to apply to Figs. 2i(b) to (d). Secondly, comparison of Figs. 21(a) to (d) (or equivalently,
44
-
TABLE I
PROBABILITY OF ERROR OF THE CONVENTIONAL EQUALIZER
WHEN APPLIED TO THE MAXIMAL DISTORTION CHANNEL OF ORDER N = 2
SNR
(db) L=3 L=7 L= 11 L=21
TABLE II
PROBABILITY OF ERROR OF THE CONVENTIONAL EQUALIZER
WHEN APPLIED TO THE MAXIMAL DISTORTION CHANNEL OF ORDER N =3
SNR
(db) L=5 L=7 L= 11 L=21
45
TABLE III
PROBABILITY OF ERROR OF THE CONVENTIONAL EQUALIZER
WHEN APPLIED TO THE MAXIMAL DISTORTION CHANNEL OF ORDER N =4
SNR
(db) L=7 L= 11 L=21 L=31
TABLE IV
PROBABILITY OF ERROR OF THE CONVENTIONAL EQUALIZER
WHEN APPLIED TO THE MAXIMAL DISTORTION CHANNEL OF ORDER N =5
SNR
(db) L= 11 L=21 L=31
46
10
a: 0
§
w o
=2-
o
m
-C
4 _
o 0
m
G
IC
0
0
O
n-
t
- IJ
4
4
0
o 0
tL
Fig. 21 (a-d). Probability of error vs SNR for conventional equalizer having L-tap TDL,
applied to maximal distortion channels of orders N = 2 to N = 5.
47
__11_11_ _I_ ^·
16
a a
o o
a
:
w
a
Ir
m
4
o
0
4
0
ca -X
a
r 0
o
a
mI- a
o
0
10
-
TDL LENGTH TDL LENGTH
(a) N 2 (b) N 3
10
0r
a
a
W
0
o
>- 10
- c
J
I-
J m
az
o
tr
(L
a-
t
10
oo
to-
Fig. 22(a-d). Probability of error vs TDL length for conventional equalizer when applied
to maximal distortion channels of orders N = 2 to N = 5.
48
Tables I to IV) shows that with decreasing SNR the performance becomes gradually independent
of the channel distortion, and the performance for all orders
of maximal distortion channels ap-
proaches that which we would obtain using a matched
filter operating with the nondispersive
N = 1 channel. In brief, then, we observe just what we would expect
at very low SNR: the per-
formance becomes essentially noise limited, and independent
of both the TDL length L and the
channel dispersion N.
With increasing SNR, we note in Fig. 21(a) that the first
observable effect regarding the
conventional equalizer performance is that due to the
channel dispersion, as the curves for all
L are no longer able to stay with the nondispersive
N = 1 curve beyond about - 5 db, but instead
exhibit an increasingly greater probability of error
relative to it. The TDL length L does not
become a noticeable factor for the lengths considered
until the SNR reaches about 2 db. Beyond
this point, the shorter TDL one-by-one exhibit performances
which cannot continue to match
that of the longer TDL with increasing SNR. We likewise see from Figs. 21(b) to (d) that these
observations hold for the higher-order maximal distortion
channels, the main differences being
that their increased dispersions cause even greater
SNR losses over the N = 1 curve. We there-
fore conclude that at low SNR the channel dispersion,
as well as the SNR, becomes a limiting
factor on the conventional equalizer performance, while
at intermediate SNR the TDL length
becomes of importance as well, and for any TDL there
is always some SNR above which we be-
gin to suffer a loss in performance due to its finite
length.
Finally, at high SNR, each of the curves in Figs. 21(a)
to (d) exhibits a limiting performance
level, which we see is strongly dependent upon the number
of TDL taps used. This effect of
TDL length is better seen through reference to Figs. 22(a)
to (d) which plot probability of error
vs L directly, for different SNR. For any given L,
we find there exists a minimum probability
of error which can be achieved, regardless of how high
the SNR becomes. These limiting values
of the probability of error depend upon the residual output
sidelobes which we incur even in the
absence of any additive noise, for any TDL of finite length.
The residual sidelobes remaining
at the conventional equalizer output are considerably larger,
under both the D and D: distor-
tion measures, than those at the decision-feedback equalizer
output, as we discuss further in
Sec. VI.
Note that the low SNR curves in Figs. 22(a) to (d) have
very small slopes, illustrating the
point we made earlier that at low SNR the length of the
TDL is relatively unimportant. With
increasing SNR, the slopes of the curves increase until
attaining that of the limiting performance
curve determined by the TDL length L. These curves
enable us to examine the advantages, if
any, of increasing the TDL length with a given SNR, or
of increasing the SNR with a given TDL
length, at different performance levels. For example,
with L = 5 in Fig. 22(a), there is little
performance advantage in a SNR increase from 20 to 26
db, while with 21 taps it can make an
improvement in probability of error by nearly two orders
of magnitude. On the other hand, for
a SNR of 14db in Fig. 22(a), if we double the TDL length
from L = 12 to L = 24, we gain a fac-
tor of only 1.16 improvement in performance, while the
same doubling of the TDL length oper-
ating at 26 db would improve our performance by a factor
of about 46. By viewing things in this
way, Figs. 22(a) to (d) can thus enable us to evaluate the
effects of variations in SNR and L un-
der different operating conditions.
Having determined the conventional equalizer performance
and obtained curves illustrating
the relationships between TDL length L and SNR for
the maximal distortion channels of different
orders, we next investigate the operation of the decision-feedback
equalizer in Sec. V, before
going on to compare the two equalizers in Sec.
VI.
49
As we anticipated earlier in Sec. III and will confirm later in this section, a decision error
by the decision-feedback equalizer can lead to error propagation, or error "bursts": the situa-
tion where the initial decision error causes additional errors on succeeding bauds with high
probability, these errors in turn cause still more errors, etc. The probability of error on any
baud, of course, depends upon both the forward-TDL output distortion and upon the "state" of
the feedback-TDL, that is, the number and location of the erroneous decisions it contains.
Lacking knowledge of the probabilities of occurrence of the different feedback-TDL states and
of the corresponding probabilities of error, we could not imake a direct performance calcula-
tion for the decision-feedback equalizer as we had done previously in Sec. IV for the conven-
tional equalizer. Instead, we determined the decision-feedback equalizer performance through
digital simulations.
The N-L cases we simulated were the same cases we considered in Sec. IV with the con-
ventional equalizer where, recall, N is the number of bauds over which the channel dispersion
extends, and L is the number of taps available for equalization. With the decision-feedback
equalizer, of course, L now represents the sum of the taps on both the forward- and feedback-
TDL filters. Our first step in each case was to solve Eqs. (94) and (97) for the forward- and
feedback-TDL tap-gain vectors, respectively, thus minimizing the total output distortion - just
as we had done previously for the conventional equalizer.
Using the sampled channel autocorrelation function appropriate to the order of the maximal
distortion channel of interest, we next found the response of the forward-TDL to a single
e0 = +1 pulse transmission. We then took these resulting output samples and used them as the
tap-gain settings on a single transversal filter, providing us with a filter whose response to an
input sample of +1 was identical with that obtained when passing the 0 = +1 impulse into the
cascade of our equivalent channel, matched filter, sampler, and forward-TDL. This is illus-
trated by the upper branches of Figs. 23(a) and (b), where we have labeled this single transversal
filter as the "equivalent signal filter."
(a)
Ck
SIGNAL _I EQUIVALENT
SAMPLES SIGNAL FILTER
(k+l
(b) ~FEEDBACK-TDL
H
)nk+1 ~ OUTPUT
NOISE EUIVALENT
SAMPLES NOISEFILTER
Fig. 23. Equivalent filters for simulating decision-feedback equalizer: (a) model of digital com-
munication using decision-feedback equalizer; (b) equivalent model used in digital simulations.
51
In the absence of previous decision errors, the determination of the probability of an error
is analogous to that presented in Sec. IV for the conventional equalizer. The output intersymbol
interference distortion for M sidelobes before the mainlobe is given by
D = q-k k (151)
k=1
52
['I
C C
CC
C C C C C .C C
C C C )C)CC C C, (..,
C)C)C))CCCCCCCC)CCCCC)CCCC
L C, C
C C C C C C C
C C' C C
C) C) C, CC,
c C,
C C C CC
C C CC
C C C C C) C
-
C C C, C'C
C.' C' C C' C CC C C
C C
C C Cr CC,
I C C, C C C
CCC)CCC)CCCC)CC)
C CC C C C) C
C),-C
C C,~~~~~~~~~~~~~~~~~-
C C, C, C C C Q C, C
IC C C.
C C C CC C,C C,
CC)CCCC)CC
C C C CC, CI
C C C CCCC' -
C
C CC C c C C C Ce O C C 11
C,C C C- C, C C C C a C C C C C C C C C C C C C C C·
CCCCCoCCC)CC)CCCCC U C C C)CCCCC)C-C)CC)C)CCC)CCC)C)CC)UCC)
CC C CC C C C C C C
C C ) C, C) C C C C C) C' C C) C' C C C'C - CC
C C C CCCC C C' C C C C, C C C
C C C C C C) C C C C C C C)C, )
C C C' C C C C C C C C -C C' C CC C C CC
C CC C, C C.C C:CC C'
CC OCC)OCCCCCCCCC
C) CC )CC ) C C C, C C C. C C C C C C C C.
C c C)C IC'CC C)C))CCC'CC)C`CCCCCCC
CCCCC CC )C CCC
C C C, C C C C C Cc C
))C C
CCCCCCC1C)CrCCCCCCC ) 'C -C C C C C: C C C
C' C
C CC C C. C. C C' O C CC C CC C C C) C C, C C C' C C C C, C'
C C C C ) C) C C
C C
C' C, CC C C CC~ C C C C C) C- C C C C C) C C C) C) C C C C C C C
C)C C C) C C) C C C C) C C C C) C C C C C C C C
' C C C C C C C C C' C) OC C C .' C) C C C C C C CC C
I-cC
C C) C C) CC CC) CI C,
C C) C)C) C
C C C C) CC C C),C
C C C C C C C- C C)C C CC C C C C CC C C CC C C) CC C C CC CC )CC CC C) ,C C C CCCC
CCC C IL)C C)C)CCC
C .CCC C C) COC C
C C
CC)CCCCCC ) C) C C) C C C C C CC C C, C ) C)C 'C ) C
C cC CcC)C) CC C C I CC C C
C'nCC, C C C C) CC)CC)C)CC
C C C C C C C C) C C C C C C) C C C C
C CCCCCCCCCC)CC
C)Co CCC) CCCC C
C C C C C) C' C )CC C) C C)C)
C) -) C)C CO 'CC CC)CC)COC C~COCCC
C C,LC)CCC)CC)C)CC)CC,)C
CL.,C C "CC C)C 1 C)CCCCCC
C 1 CC C1
C)C C)C,
- ) C C C)C.-~CCCCCC)C)CCC)
CC) C)-CO
OV
OOOOOUO . 00 OUU 00CUUUC
UO
CC C C CL L,~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
C)C)
CC) LC)C) C C C) C) C'C
C, C C
OUCUO O
C)
CCC)C
C
-C)C)C C) ) C C C C C C) C C C C C) C C) C - C C -
CC)
C C C CC
)C CC) )C C )C
C C CC)C) C) CCC)C C)
C C CCC C) C C- C C)
C)D.
C C C C) C C ) C CC)CI'
)C)C
C C) C'C) C) ) ) C C C C
) CI C C ) C) CC C C C C C C) C C) C C)C C C
CCI CCC)C
C C C) C I L"C
C) I)- C)
LIC
C).
) ) C
C ) CC)
~
C2 C) C C C C C C) C, CCC
) CCI CCC C C) C CC
CCCC)C)CCCC)OC)C'C))CCCC-C2CC))C'CC)CC))CC)CC))CCCC'C)CC C C C C) C C C' C
C CC) C)C C CC C
C C C C C C
C, C C C C I O C C C C C C C) C C , C - C C C C C
CC C C CC CI C
CC C C C C C C C) C) C C CI C C CD C C C)
) C C C C C C C C C C C C C C C C) C C)
C) C C C C C C L) C C C C CC ) C) ) C
C1 C) C C C C L' C, C CC C
C L)CC)C C C C
) C. C C C) C) ) C C) C) C
C C)C
C C)C)CCC)
C C C C C) C C C)C C) C CC C)1C) ('C--CC C) C) CI) C) C) C1 C) C) C1 C)1 C C) CI)C CCCC)C
CC)C C C)CC CCCCCC
C C C C) C C)CCCC)C , C. C) C. C)~ CC)I
C C C C) C)CC CI)CC)L,C))CC C)CC C)CC C) )C
C)c C ) C C, CC, C C
C C)C )CCC CC C )C
C C C)C C)
C C) C C C C C) C) C C C) C C C) ) )C
C)CCCCC)C)C)C)C)CCC
C C)CCCCC)CCC)C)CC)CCCCCCCC)CC CL
)
C'C)CCCCC LI C) C) C) C LC
C C I)
C) C)C
1 C
1C.
CCCC
C)CCC))C)CC)CCCC) C I)))C)C C, C-C, C C) C) CC)C) C) C)
C.)CCCCCC
LC
c.~~~~~~~ o1 11 11 11 ~~ ~ ~ ~ ~ CC )C. ,
o () I.C
g,IL,
~ ~ ~ ~ ~~
u_(_~C, o I CCC'
o;I.
L., u,
C,
(.~
u
I
c7, I -U.
.
)gf;o
(.IJ1
0 o¢ , L,o1 U I I. ~~ ~ ~
C) C ) C C) C ) C ) C) C.'UII
.I C C)- C C) I, C) C C) C C C) 1C Cc,,, 01 C) C '0 C) ) C C)1C C) C C C C C) C) C C) C .C, ) C C ~ ~ ~
L
L
1 1
~ ~ ~~
C)- -
C C)C) C)C)CCC)C' o~~~~~~~~~~~~~c
) )CC.C)CCC
, CC)CC)))CCC))C)C'
L. CC)
C CC
: C C 0 CC )CC)C ,'J C C CC C. ) C C C) )C C) C )CC)C C, .- ) C) I
C' C U C)C)C)C)C)(JC)C)CL'C)C)C
i iC C)
) C C CC C C C C CCC C C )C C CC C
C ) C'C U C C
C;~~~~~~~~~C
C)C)CC CC C' C F Cc
C C)C C U C UVc' (IC C C
CoCc)ccCC,CCC)C)))C)CCCCCCCCCCCCC)C
C) C)- C,)Ccc,
10.1 C)C
CC CC-'CCC).j,
C, oC-C -CC)CCCCC)CCC,. , C CC, *C C
CCjj
-CCC)CCCCCC)-)CCCC)C)')CC))C)CCC)CCCCCC)C.""CCCCCC)C)
CC)C CC---CC -- C'C)CC CC)CC.---'.--C'C)C)C C, C ) C)C)C QCCC CC,
C)C)-.-.CC C)CC'C)CC)CC C-C C'C C CC) ~,
ICCCCC
C CCCCC
- CC,-'CC C))CC
c2
CC.CCCCCC)CC)C)CCC C)CCC)CCO C C)CC)CC CCC)CCC CCC-CCCCCCC)'C 0..
~c)C)
c-CC
C - CCCCCC
cC).-.CC CCC)C)CCC)-CC
C)CCCCCC-C)CCC-CC:CC)))'-CC)C)CCCCC)CCCC
C C CC
u~~~~~~~~~~~~~~~
CCC C
CCCCCC(.CC)-C.-COC)CCCCCC)nCCCCCc'CC.Cc)
-)
U
u) C . CC
C)- U
C.
,)
. .U gI .UIfC. {C,
C.C
CC.g
l r.g
C UC
g g.Uc.jIU(_ ~, g U u /~, C.,CC)C
C,-'CCCCC)C C) C)
C C C
C;~~~~ cCI 0CC C JU{J
LC' c C' C -
U C
U C u C /.
C C'
. f~ /gC g) og
.... L C C C C C)
g ~J
CCCLC,)CCC_
u g { C.
UU k.; C_, g
C)
u C C C) C) C
g' g k-Uu .
U g (
g (g U
CC
cCC),,)CC
/g gc g (J
CCC C
,
g
C
~ CC C
C..CCCC)C
)U'
_ CCC
gg, U
CCC C
) C' C C
-u
..
C)C
C-
00CC-CC .-. CCCC)C)-C:C)CC
CCC-)C;C-C C'C, .. U
~
C C U - 'C--CCQ00
CCCOCCCCCC)C)CCCCC-CCC-CCCCC'-----))'CCCCC)CC)
Uuu k C2C)CC-CCC.-.CC)CC)CC'C)Cj~
~. . .U . U . . U J U
C
'U~k.
CC C)-C CC
U (_,
_ CCC~
u J U U
C''.'C'C
IC
u U
. .' C
C
. .-
C )C CC
C)C,
CC
U )CC C,
{g ( Ug
C'Cuuu-.
CC)C
.CCC
0CU
uC F CC
U
{.
. ( g
' -CC v
C L)
u U u U ~-C u U uUuUC
C)CCC)C)CCc)C'-)))CC
~ )U
u u r_ U_ u U C- uj ~ U
)CCCCC~~CCCCCC,-C)CCCC--CC-C)C-C
C-C
%C C CCu C C
UU~,
)C)C-COC
UUU UC CCC
C CCC CC U ' C) ' CCCCC)C
: U C C CUV C C C C- L~))
UU..U C '~Cj
U ClU-UC U %-CC
U C'.'C'C
U
C' CC(
CC U
U UC-C,
U u- ~.~ U
)L-~~ '
U. UUU
· CC,
~
CCC U U U CC)rIC-)CC
C C)CU C Lu U U
U U~ C UU U' CU C
~c'.--~ ~ C.- OC C
C O O O
U U,C)C
U ) U UU(-}~
C C - C,
:U
C u
~
C C O~C
L, C, C,
c
CO
O C-
C - `
O C:C
OO O O C C C ' ~
C ~ O I c
O
O
O c
C
C)CI'.C)C)CCCC)C.-C)C U O C OU
( OO
- , U
OOCUOCOUO 0 OO
C U V C C~
U C
C)u
'J~
L:
)u--
~ ~ ~ ~ ~~~~.
C)C
C)C
-
)
I,'-C C)g,
( CI ,. CT C~ ¢ C UiV(
C
- C, C3 C'~~(
UOCC;LOC)CCL)'LC;,C
C ~ C.U C'i'
C CC
C
C
C
uuuu~
C_
)UC'(
C C , UL'J
u u
C.C'U
) L.Cu, C )' C: )
_,
,'')-UU'U)CUU"UCUCUCU-.CU-U
LCuC .' L.
C C ).
u , L' ' U -U ' CCU 'C U U
uuu~ ~
C'C,))C')C))CC'CC.)CCCCC'J',,'CC)C-C)''C)')CC'C-'CCCC)UUC
C)U
U U C'L ` U U U U U U U 'U CC k U U CC C'
I
C)C)C'
C)CC)"CCC
CCCCCCCL'CCCC'C)CC)C)CCC)L'C)C'CCC.L'C'C UU)CU UU C)CCC)C
'~·( C-C~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Q
, uCC C).- CCCCC
U CC u Cu
) u C 'C - CC)CC)CCCCCCI
C C %'I C,)I)L
C C
C)-C)C) . C CC)CC-
C. - u C - C
C ' C)
53
C C - C. C C` C C C C C C. C C C C C C C C. C CC ,
C.CCCCC
C I C. C CC C' C C. C'C -
C. CC C C
C. C. CC' C
C C CCC C C. C CC C C.CCCC CCCCCCCCCC'.C
'C C ) C C C ) C, C C C C C 'C C
C C C C C C C ,C )C,C C C C C C C C ,C C .
C C .) C C C. C ) C ) C - C C C C .C C C C C C C C C C CC C C.C-
C C C' C C C C . C )C C I'C C C C C C C C
C
C C) C) C C C
CC C.' CC)CC))
C CC,
C C)CC
C )C) C
C CCC C)
C) CC) C, CC C C'C C. C.C C-C C C! C
C C.C CC C C CC C O)
C. C C
C. 0
C
c C C. C C 0 C 0
C. c C
C C C) C)
C CC.
C C CC CC)
C CC C . , C C
. C
CC C C C C,C C C C C) C .C c
C. C C.
C. C c CC C C C C .)c
C C .C. C' Ci C C. 0 C C C
C, C . C-, )C . CC ) CC.CcC C C. C
o:
C
C,
) )
C C).C )C CC
CCC
CC.,CCCC)C
)C C)
C
,C)CC. CC C
C
C )C)
C 'CIC)C
C~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
C 'C. )
C C )C
C
C C'C
C C
)
.CCCC
C)L'C . C)
)i C
C CC
CC,)CC
.
C) C
C CC
)C)CC)LI
C )C )
C) C.
0 'C
C)C
C.
C C .'
C C) C C CC
)C)C
C CC ' C
C C CC
CCCC)
C CC
C C C C C C
~C
CC C CCC
CC,
C C C
C C
CC
C C
CC
CC
C-
C C C ,
) C
CC
OC C, C CC C
CCO
.)
C C) C C C C C, C C C C
C C C C C I ) C) C0 C)c C) C C
C. C C, C C C CC
C )C CCC
OC
LI'
C
CC
0) ) C)U
C CCC )C.
oc
C)C C)C Uc
)c6' CC
C cc
C C C ' C ) C C. C , C C
cc ) c(~c o o)
CC.
c) ) uco
c.~cc ~
~
) ) C u0I
C C)C
C C CC C~
C, C C C
C ~ )C
C)
c ))CC
S:
C' C C - ,C
. C)' C.- C¢I0
)' F,. )C )C )C C
C )CCc,
o
C C
Z
L
o
CC c L
,
oor~
UV
C
c
c
L; C, LI 11Z -, c)
C
C c C C)c)
L, 1C C
LI
UU~~C
I , 11 C C;C- C,
C C
C, C L, LI Qo C C,0 I.;C)CI- CCC, CCC
CC C I.;CC) C CC C
1- C-
cc: 11 11C., 0
C
, c
C
ooCC oc
C CC , 1,
C) C C))
C C C) C)C) C) C) C C C CIC
C C)C)oC))C)CC) C', -CC CC '
C C) C)C C
C.
) C ) C C C,
C)
C ))oLC -, o.) ) CC o
C
C'CC C C) C,.C C C
C) C CC) I
fl C) C)C) , C) ID)C,
Cc,C C , C C) ,_oCCC) C.
C) C)C))CCC'C)C)C)C)C)C)')
L, 1 1 1
C)))CC'
C
CC)'C)C..C)C.)
- uc
LI
C)C)CCCCC C)C)C
~~~~~~~~~~~-
1 LI
-- L-'
U)
I
· C)C)C
-" -
).
UC'C C) C)
1 Ir ) I.: Q L I.,
C
i11 V
)C
CC CCCCCC)))
CC)C)C)C)C).C))),
C)
I '
C
C))C)C)C)C)C)C
U L' I,
, f V C:
u~~~~~~~ C)
r--. ) CCCCC
C CC) UC)C))C)
U U V
U C)LIC)U
C U
C) C).
V
C
C
)))))
C)CL
C)C'
U
LI
. C'))
, )C ),C,
CU
)C
00I
a
C C )C C) C) Z~r_;
C )C)CLI uuuuuu U
C)CCC))
C.C)CC)CC)ICC))C))C ulu
C)(.)C)C)'C)C)C)C))C)CC))C) uuuuuuvuuuucuuuluu
uu
C )CC C)CUC,,
IU ' 0CCC~
I
C
,3C
C' C
.,
,.C, C
CCCC
~~~~, LI
C
U
,.C:
,_, )CCCCCCC
, LC I U
(-;)C)C)CC))CCC)CCC
)C)C)C)C
,
,C0
C,0
o L C)
'C
Q C)C)C))C)C))C)C))C)C)CC)C)C)CL)
¢., , C_,
C,
,I,L.;L
,UU
7,.. 0 C,C; r ,_ Cr.
II
..
C C,,.,..,.0c ) c_
C..
u (I.;
¢,,j ,r.~ u t_
') )C)C)CCICC
)) CC) C) CCC CC)C)C) C C, C CCCC C CC
C CCC
C C)C)C;CCC C C t)
II rC))L CCCC CCC)CCC
C~~ ~
C CC)I
C,
C)CC: )C'
C,
C C'
C-C
C,C `)C
C)C C)C CC))C I
C)CC))CC
u~~~~~~~~~~~~~~~~~~~~~~~~~C
, ,) , ~~~~~~~
~ ~ ~ C)
C I C
C)
C)C) C,C
uu..J
C)))
. ...
C))
CC)C))C)C)C
C, C-C)C)C),
. .. C C:,',,0C C' C))))))CCCCC
)CC)C)
C
u
O C) C
c
cJ
C CC
,,... 0
C· - C
C
CCI I-
CC c; C
C
CuCCC
C
C
C~
CO
1-1 1)
CCI,
*CC)C
C) CC
C
V
)
~I
~
OCC
IM)
CM) C C
C
C
C, C
C.
)M CC. CC C
'
C.
C
C
'O
C
C I
C
C
CC).CC
.L
C
C
C C'C C
CCM
E)tIll.
IC~
1 CC
)
I) ) )
C CC CC
CCCIM
CC)
CCCCC)CC C
C
C7
C, C
C
MC C C
C
C
-,
C Cco
C CC
CC
C-
C C,
CC"
c
CCCOO
)
C ,C
C I'U
C C C
C CI1 C
" C cc, C) (D C CC
C
CCC C CCC
CC
C, CCC)
CC
C1
C -CCC
CC I UCC CC
C fl
,
C, C CCC
C....
I~~
1 1
CCCCCCC
C
C,.
C.,
C
C'
C)
C C.C
flC CC ' CUC .
LC-~~~~ .
CL CC C )
..
U
0
liii
C CC.........
EMC
C.-1
C
... )
CC
CC.cC)CC.C
,
CC)
C
CCCC
C)
C .C
CC
C.
,
C).
C., CCCC CC CC
C) C
C
C
C
C
UUCUUUVUUUUUUU~~~~~~~~UUUO
CCC'
CC
C
C,
~~lC.
cc)C
- -C
C
C
I OC CC C) c )CC C C
CC)OC C
C
C CC cC
C
C ... .
C Iu
C
I-, u.. (ICL;C)C'C.C)C)C.C.C,
C
"
C C
C
...
.....
I
C)CC
C ' C
C',
.. C.C)
CC C
C
CuCCC
) C
C.CC,
cC"CC
C
0C
C-. C CC
C C. C-C
C
. .. C)CC)
C C
c CC,~~~~~~~~~~~~~~~Q
CCCC 'CCC C.
C..CC)...C.......C.C
CC . . C .,CCC
C )C C )C C-,)
" ).CCCC)
C
CC
C cC
Cu
CC.C
CCCC
CC.
c
M
C)
clCCC
C.ICCCC
C C C
UM C
C
C.
C)c. oC.
) M CC)
CcC CCCc)
CD C.C)I-
.CCC.U
.c C CICC,
oc)
oC'
CCCO
CCCCL))CC
C.1
CCCCCCCC,,, U .
CCCC
C C .
I
cC)
C C C
I ,I-C C
CC.CC
C
U
C C
C)) C
U g U
C . C.C......
C C C C)
g U
-o
.
.0
0
...
C fCC C
LCC I, CC
M M C, C) mo r CCC C C. CI'
CC.LC
·
I'lCC
c cC Cc o
C CC CI.C
) 1)-
CCCCM,C.C
C C C) C
U
C.))C.I C
. C, C'CIIILC CCC CC M
QCCC
CC CC C L, C I)--.
C'CCCC CC,
MM C.
CCC) C C C)C)
M IC
"C C CC)
C C C.)
CCICCCC
C, C C
C
CCCCC)
CC CC)C C C
M
CCC
C CC
)C .
, Q ,CC 0
C) I UC C
CCC); C C- LC' C CCC ' C , C C LI C C, , Cc
cccn~~
c jC" c "U o o CM
o§8 ~
C L LCC))CC L C CCCC)
M C.'C)(.: Q: C'U L C r.
CC
CCE
CCC
OCC)C)CC)C;~~
CCC) C)CCC)
C
UCCC)M C)·
1C))C HEM
C 1 )C)CC
CC c C)C.uCCC
C
Cc uuC)C)_
ulccco
OCC)
1CC C) C
oc
'~ C oC:C
CCCCCC)
e C)'I c'~·o
cc
· C 0CC)
CC~Ccouoc C)L0 °oc
~ -,oc Q
C C).o
L' C'
C
rJ
C: C-U.21 C ',3
0 ~.
II C:
CC)UCC
c- .:I
, CC
I CD L~ ~ ~ ~~
L
L~
C CC
rU
, C.'
. C
C
II
CC)
ID c.
I.
CCC
C)C
CO
C
-,C
- ~,-c~e, .
1,
0 C
CCCC)CC
'.¢¢
C
I_¢DO
C
)C
.0 . )C)
~
.L. O i
M iC
c~O
CC
iF,CI O
CC ~CC)O~~i
C
C onf C C
C -I
CC)C
C
C)C C)C ) c )C C C C CcCC C C M
~
~~~ ~
C CC LI C) C)CL C)C MCC C CC oo C)C
C)C)C c c~oc
C)C.I-
ulo '
U CC ) u.C
cc u L,
' I) CC c.
o iCOLO C c
¢ C
~ )I C * C
C)C I
,
o L~
c._
L' I u '~
,
C'UUU~
c I
c-'
~ 'I- ~
~~~~~~-~
UIUU~~~~~~~~~~~~~~
Oc~¢
c
, c(. ._ c., c.- ICc_.
~ I Lo Ci
~ cU~-cc
UU -11o ~
C;,
C
LI~c,~
uU -
C ))~
LI U
o,'
--"I
L I i. U
IO.;CC
I I:CUUO
MCCC
c1
c- u ~ I
~ c -i ~1u-c
LI; OOCC)
LI~
o
- ~ o LIc ;11c , o
CC)C
C M: CC
C)CCC UOCOO U cC
i c, LI CC) o
CO CCC , C C , I u I ~,Ioo IoouI¢
uuu uU~uIIUllU
uu
L.,L,
u uu
IL-''-
u ~_~
I '
~u ,;u -uu. ~
C))C CC~
1 '- ,.
1
)C
C,
'
c., .~
C
~u
uu
u
-'~L'c
C
.....
..
~~
))
~I
,
C))
I
-
111
I
LI
u~~uu~~cDr
I
UI .......
C C
I c:
u
I ) c
U u iC)Ii, .
) 'll
r.
LI
_
I
III
,
~~,-u
L,
,,,
',, CO
~,'
1 ou
~
CC)C,
,
1CC
L
CCu
u~~u~.,,r
CCCM-
CCIC) I
..
C
CC)C)
,-I
..
c,~
uuu uu uu~ m,; CC).
*C)I u . ). , ,,uuuu uu ou uu
u.u,, , u
i u
~ou
,o o
,:,~ ,
-ouc u iu
uo
cc i
c:.c-ouo·~c,
C~~ijC
u~ u.. u
UU
uu,.'u.
O
uiu . u
C CCC)O
r.u u u
OO
~u C)C
O C).)
uu-
C)
uu i
u)C)C
uuiu ii o~iu
CC))C
u uL.
uiu
ou i ui u o uuiu ul c
uu CCC) iuC C ,C CC
C)C- C. u CC u uu)u C uu
Cu)uu) uuuiuuuuuuuu u i
CC C..C CC C) C C )
IUy. Uu uV ~- uUIVU UC ,,-
I , V~uiu.
IU iu~iuui.
U. i U ;uu~
VV C,. V) O
u
54
I
which differs from the corresponding distortion at the conventional equalizer output, as given
by Eq.(121), in that the distortion arising from the k for k < 0 has been eliminated through the
decision feedback. Equation (122) applied to the probability of an initial error occurring, is
still valid:
2
where we recall that a is the output noise variance. Analogous to our earlier derivation of
Sec. IV, we define
and b b ... , b M to be the ordered version ofb b... b, so that Eq. (134) still holds.
'
Defining q 1, 2' .. 77M to be the corresponding rearrangement of the 1, 2 ... l' the out-
put distortion in the absence of decision errors becomes
M
D= bk77k (154)
k=
which is formally identical with Eq. (135). The difference now, of course, is that the 7 k are
independent binary random variables, with k = +1 with equal probability for all k. If we de-
note the occurrence of an initial error by e i , Eq.(136) becomes:
177~~~
except that Eq.(155) involves 2 M terms, compared with the 3M terms of Eq.(136). Since this is
still an impractical number of terms to sum over directly, we may proceed as in Sec. IV, where
Eqs. (137) through (146) still apply, to arrive at upper and lower bounds on the probability of an
initial error, given by
Thus, by modifying appropriately the error tree algorithm derived in Sec. IV, we may compute
the upper and lower bounds of Eq.(156) to any desired closeness to one another. The descrip-
tion of the modified algorithm is nearly identical with that given earlier, the only differences
in its implementation being: (1) the number of paths leading to the right from each node is now
T
only two, corresponding to nk which are binary rather than ternary random variables, thus re-
ducing the "flagging" requirements at each node, and (2) the weight W(K) of Eq. (143) now becomes
W(K) = 2 -K (157)
which is independent of the particular r7* sequence taken (i.e., the particular path selected in
searching the tree), and only depends upon K, the distance between the present and the first
node. This means the sequential computation of W(K) as given by Eq. (147) is no longer neces-
sary. These two differences thus result in a somewhat more simplified implementation of the
algorithm than we incurred previously for the conventional equalizer.
55
th
ni = number of errors occurring in i burst.
13-42-10173J
no n2
11111I
111I 111111111 11
I go ---. C, --- 02 -- C
2
Note that, except for the 2-L cases, c i and ni may differ, since not all decisions occurring
during a burst of errors are necessarily incorrect - as we can observe from Fig. 24(a), for ex-
ample, in the 5-21 case. To estimate the overall probability of error, we want to compute the
ratio of the number of bauds in error to the total number of bauds, as given by
En.
p(error) = (158)
Zgi + ci
as the number of error bursts encountered, P, becomes large. Normalizing by dividing through
the numerator and denominator of Eq.(158) by P, we find that
B
p(error)= e (159)
S +
g r
Now, the mean guard space is simply the reciprocal of the probability of an initial error
S 1 (163)
g P[ei]
56
· __. ___
which we are able to determine to any desired accuracy using Eq.(156) implemented by our
modified error tree algorithm. Next, we may write the mean recovery time after an. initial
decision error as
=
Tr [Trle i , -] p[7le i] (i64)
77
and, since
p[eil71] p(7)
p[1 lei]= p(ei) (165)
where, just as with Eqs. (145) and (146), the first summation is to be taken over all distinct 77*.
Now, for :77E7,
Tr Z [TrIei,Tl]2 (168)
7*
77E
and note that Eq.(141) applied to the present problem renders bounds on the probability of an
initial error for any 77E77 *:
then we may substitute Eqs. (167) through (169) in Eq. (166) to obtain the bounds on the mean re-
covery time
Using Eq.(170), we could obtain arbitrarily close upper bounds on the mean recovery time, if
2M-K
only we knew the quantities T defined by Eq.(168). But, for each 77:-4there are 2 different
q7Er'77,thus Eq. (168) represents the mean recovery time averaged over the equiprobable r1E77*.
Hence, for a given _*, we want to take the various 7E7'* and determine the mean recovery time
after an initial error has taken place. One way in which we may do this is through quasi-
simulations, which we now describe.
We randomly select an 77E7*- and put it into the equivalent signal filter, "freezing" the r/k
in their respective positions. With no errors in the feedback-TDL, we then pass a sequence of
samples into the equivalent noise filter of Fig. 23(b) until a decision error occurs. After this
57
- ~ ~~ ~
- -
initial error, we then "unfreeze" the sequence within the equivalent signal filter of Fig. 23(b)
by permitting new signals to enter it, and normal simulation continues until the error burst is
over. Through performing a sufficient number of these quasi-simulations, we may obtain the
conditional recovery time of Eq.(168) as accurately as desired. Further, by doing this for each
distinct 77'", we may weight these expressions of Eq. (168) to obtain the bounds of Eq. (170) as
closely as desired.
All this is conveniently achieved through further modification of our error tree algorithm.
Recalling our previous discussion of its implementation in Sec. IV, the major difference now is
that, in arriving at a node where the bounds of Eq.(167) are sufficiently close, we then perform
the quasi-simulations described above for the current value of 77*. The mean recovery times
of Eq. (168) are weighted appropriately by PL(K) W(K) and PU(K) W(K) and accumulated, to form
the lower and upper bounds of Eq. (170), respectively, as we search through the tree. Simulta-
neously, of course, we accumulate the bounds of Eq. (156), thus obtaining bounds on the probability
of an initial error (and then to the mean guard space) via Eq. (163).
It is clear that we may obtain bounds on the mean number of errors per burst in exactly the
same way as we have just shown for the mean recovery time. Again, the computations fit con-
veniently into our modified error tree algorithm, and we must further determine only the average
"
number of errors per burst occurring for our quasi-simulations for each 77- . Denoting this
average as Be we may then use an expression identical with Eq.(i70), except with Tr and T
replaced by B e and Be , respectively, to obtain the desired bounds on the mean number of errors
per burst. Similarly, we can determine bounds on the mean square recovery time, which we did
in order to study the variance of the burst lengths, as we discuss later in this Section.
The rationale behind all this is as follows. The difficulty with high SNR simulations, as
we noted earlier, is the low error rates we encounter. These, in turn, are due to the small dis-
tortions appearing at both the equivalent signal filter and equivalent noise filter outputs. In par-
ticular, at high SNR both D: and the output noise variance a are generally sufficiently small
that most of the time neither the intersymbol interference nor the additive noise acting alone can
cause a decision error. This is strictly true for the intersymbol interference acting alone when-
ever D9 < 1, which it always becomes at high SNR with the decision-feedback equalizer, as we
will see further below. Similarly, the probability of noise causing an error in the absence of
intersymbol interference is negligible compared with its probability of causing an error in the
presence of intersymbol interference opposed to a correct decision. Thus, at high SNR, errors
occur most of the time only through a collaboration of intersymbol interference and noise dis-
tortion. All this, of course, is true only in the absence of decision errors, for with decision
errors in the feedback-TDL, D can easily exceed unity and error propagation can occur due to
intersymbol interference alone, a matter we describe in more detail below. Thus, in the ab-
sence of previous decision errors, an error in our simulation involves the simultaneous occur-
rence of a "harmful" intersymbol interference distortion and a "large" noise distortion, which,
both being low probability events, means such errors occur infrequently. The quasi-simulation
approach described above effectively determines those intersymbol interference sidelobe com-
binations which are most strongly opposed to a correct decision, through the efficient selection
of 7 via the error tree algorithm. The noise is then permitted to cause an error, which even-
tually occurs. However, we must have the noise correlation correct at the time of the initial
error, so, in general, the possibility of "inserting an error" is not a valid approach. At very
58
high SNR, however, as we will observe below, the noise distortion becomes virtually uncor-
related; thus, inserting an error at high SNR is a legitimate approach. Both methods were
adopted, where appropriate, in our quasi-simulations to determine the decision-feedback equal-
izer performance at high SNR using Eq.(159). At low SNR, of course, normal simulations were
performed, as explained earlier.
We present the resulting error-rate data in Tables V to VIII for the maximal distortion
channels of orders N = 2 to N = 5, respectively. The burst and guard-space data we obtained
will be presented and discussed further below. In Figs. 27(a) to (d), we plot this error-rate data
vs SNR for various L, where L is the total number of taps on both the forward- and feedback-
TDL, with a separate figure for each of the maximal distortion channels considered.
At very low SNR, the performance becomes noise-limited, independent of both the channel
distortion and the number of taps used. This is seen to be true in Fig. 27(a) for SNR less than
about -10 db. Beyond this point, we note that the error rate curves gradually become worse
than the nondispersive N = 1 bound. The effective loss in SNR for a given error rate is seen to
reach a limiting value with increasing SNR, however, becoming approximately 4 db with L = 3,
and 3db for L = 7 at an error rate of 10 . To further investigate this matter, we greatly com-
pressed the vertical scales, enabling us to plot the 2-L error rates down to 10 -35 , as shown
in Fig. 28 where we observe that the effective loss in SNR does become constant with respect to
the N = 1 bound at high SNR, being about 2.8 db, for example, in the 2-21 case. We see that
these observations also apply to Figs. 27(b) to (d), except that the loss in SNR is seen to become
greater as the channel dispersion becomes larger.
There are two major reasons for the degradation in SNR with the decision-feedback equal-
izer operating at high SNR. For any finite length of the forward-TDL, with increasing SNR its
tap gains (which are set to minimize the total output distortion energy) approach limiting values
which are determined solely by the intersymbol interference and the length of the forward-TDL,
independent of the noise. These limiting tap-gain settings result in (1) a limiting noise enhance-
ment effect, and (2) a limiting set of output sidelobes which give rise to the intersymbol inter-
ference (these two factors are shown in Figs. 29 and 30). In Fig. 29, we have plotted the degrada
tion in SNR due to noise enhancement vs the forward-TDL length for the maximal distortion
channels of orders N = 2 to N = 5, where we recall that the forward-TDL has a length given by
L - N + 1i, since N - 1 of the total of L taps are required on the feedback-TDL. In Fig. 30, we
have plotted the limiting value of the intersymbol interference distortion measure D: with in-
creasing SNR, again vs the length of the forward-TDL.
The intersymbol interference distortion remains nonzero, of course, for any forward-TDL
of finite length, although from Fig. 30 it appears to approach zero asymptotically with increasing
forward-TDL length. As we observed above from Fig. 28 for the 2-21 case, the loss in SNR
compared with the nondispersive N = 1 curve is about 2.8 db. However, from Fig. 29 we find for
this 2-21 case, where the forward-TDL has 20 taps, that about 2.6 db of this loss is directly at-
tributable to the noise enhancement, with the remaining loss caused by the residual intersymbol
interference distortion which has a D = 0.065 as we find from Fig. 30 for this case. Thus, for
a decision-feedback equalizer having a long forward-TDL, the loss in SNR it suffers compared
with the nondispersive N = 1 channel performance is essentially determined by the noise en-
hancement of the forward-TDL at high SNR. Throughout this discussion, of course, we have
59
SNR
(db) L=3 L=7 L= 11 L=21
32 .
TABLE VI
DECISION-FEEDBACK EQUALIZER ERROR RATE WHEN APPLIED
TO THE MAXIMAL DISTORTION CHANNEL OF ORDER N = 3
SNR
(db) L=5 L=7 L = 11 L=21
32 2.51E-59 - - -
60
TABLE VII
DECISION-FEEDBACK EQUALIZER ERROR RATE WHEN APPLIED
TO THE MAXIMAL DISTORTION CHANNEL OF ORDER N = 4
SNR
(db) L= 7 L = 11 L = 21 L = 31
38 .
TABLE VIII
DECISION-FEEDBACK EQUALIZER ERROR RATE WHEN APPLIED
TO THE MAXIMAL DISTORTION CHANNEL OF ORDER N = 5
SNR
(db) L = 11 L = 21 L = 31
14 2. 49E-02 1.45E-02 -
20 3.36E-04 6.72E-06 -
26 7. 13E-09 4.67E-15 -
32 6.88E-25 1.46E-45 -
61
- -- I
W
"I
0 0
w
11
(a) N 2 (b)N = 3
I.
X
1C
I M
a:
'r t
0
(c) N = 4 (d)N = 5
Fig. 27(a-d). Error rate vs SNR for decision-feedback equalizer applied to maximal
distortion channels of orders N = 2 to N = 5, for L taps.
62
w
I-
hm
distortion channel. n
SNR (db)
N =5
z
4 4
0
a
I
w
o
UJ
z I
3
2
C,
a:
I I I I I
0 5 10 15 20 25 30
FORWARD-TDL LENGTH FORWARD-TDL LENGTH
Fig. 29. Noise enhancement vs forward-TDL Fig. 30. Limiting distortion Dp vs forward-TDL
length in decision-feedback equalizer, applied length in decision-feedback equalizer, applied
to maximal distortion channels of order N. to maximal distortion channels of order N.
63
..-__Llpl
.....
I---- 1_1( 11--I__ Y
-11-·-^1 C1^
considered Dig to be measured in the absence of decision errors. When such errors occur, the
D: become large, and bursts result even at high SNR, as indicated in Figs. 25(a) and (b) for the
5-21 case. Such bursts, however, merely increase the error rate beyond the probability of an
initial error by a factor equaling the mean number of errors per burst. This is true because
in Eq. (159) the mean recovery time becomes negligible compared with the mean guard space as
we will see below, and the mean guard space and the probability of an initial error are directly
related via Eq. (163). As we consider next, the burst and guard-space data obtained from our
simulations and quasi-simulations using the modified error tree algorithm are such that, given
the steepness of the error-rate curves in Figs. 27 and 28 at high SNR, the further degradation
due to bursting effects becomes negligible compared with the losses arising from the forward-
TDL noise enhancement.
In the course of our normal simulations at low SNR, we obtained error-burst data, recording
the relative frequency of occurrence of the various burst lengths and then computing the average
burst duration, or mean recovery time Tr as we referred to it above. We also computed the re-
covery time variance from these data. In our high SNR quasi-simulations, of course, the Tr
was necessary for our evaluation of Eq. (159) in order to obtain the overall error rate of the
decision-feedback equalizer, and we further recorded the recovery time variance. Typical mean
recovery time data are shown in Figs. 31(a) and (b) where we have plotted the mean recovery time
in bauds vs SNR for the N-11 and N-21 cases, respectively.
w
r- F-
,
W
o
0
>- 0
c)
w
w
z z
ha
2
Fig. 31(a-b). Mean recovery time vs SNR after an initial decision error with 11- and 1-tap
decision-feedback equalizers, applied to maximal distortion channels.
64
___ _ _ _ _ _
As we observed earlier, at very low SNR the noise dominates both the channel dispersion
and the length of equalizer being employed. Correspondingly, even after a decision error has
been made, the length of time until recovery is essentially determined by the additive noise,
asymptotically with decreasing SNR. As a result, the mean recovery time approaches asymp-
totically the mean time required before the receiver makes enough correct "guesses" that the
feedback-TDL once again contains no errors. The average time to recovery through guessing
with an N -order channel may be handled using the simple Markov chain flow diagram of
Fig. 32(b) where we have let the "state" of the feedback-TDL correspond to the number of the
rightmost tap position containing an error. Thus, referring to Fig. 32(a), an initial decision
error becomes positioned at the first tap, corresponding to state 1 of Fig. 32(b). If the next
decision is correct, the initial error shifts to tap 2, or state 2 of Fig. 32(b), etc. Since the
feedback-TDL has N - 1 taps, then clearly upon reaching state N, recovery is complete. In
th
Fig. 32(b), the transitions from the K state are, on a "guessing" basis, equally likely to re-
sult in states 1 or K + 1 on the next decision, as indicated by the factor of 1/2 on each transi-
tion. The z appearing on each transition represents a unit delay, corresponding to a decision
interval, the baud period. If we let Pi be the probability of recovering after exactly i bauds,
then defining the z-transform of the P. P2, . . sequence as
oo
P(z) = pi z i (171)
i=l
it follows that
oo
which is precisely the mean recovery time we desire. It is a straightforward exercise in flow-
chart manipulation to find that for Fig. 32(b)
1 N-I i
(1z) (1- 2Z)
P(z) 1 N (173)
1 -z + (z)
N-i 3 2 1 [3-42-101131
FEEDBACK-TDL DECISION
IFEEDBACK
(a)
INITIAL RECOVERY
ERROR ( Z 2Z Z Z 2
65
- _ dP(z) = 2 N 2 (174)
r dz =
is the mean recovery time we would encounter through guessing. We can see from Figs. 31(a)
and (b) that the decision-feedback equalizer mean recovery time does indeed approach asymp-
totically the values given by Eq. (174) for the different order channels, as the N = 2 channel T
approaches 4 bauds, the N = 3 channel Tr approaches 6 bauds, etc., with decreasing SNR.
With increasing SNR, the mean recovery time decreases up until about 2 db, as we observe
from Figs. 31(a) and (b), mainly because the additive noise is decreasing; hence, the decisions
are becoming better than those obtained through guessing. In addition, however, there is a
compensation effect which we now describe. The occurrence of an initial error generally re-
quires that the additive noise be opposed to a correct decision. That is, for the decision-
feedback equalizer to make an error in determining 0, say, when 50 in fact is +1, requires
in general that the output of the equivalent noise filter of Fig. 23(b) be negative. On the other
hand, on the succeeding decision this error results in an additional sidelobe of 2qt40 because,
referring to Fig. 10, the feedback-TDL now enhances the q sidelobe appearing at the forward-
TDL output rather than subtracting it out. Because q is positive for the maximal distortion
channels, this enhanced sidelobe is seen to oppose the noise on the succeeding decision, since
the noise correlation properties are such that the negative noise which caused the initial error
is still likely to be negative on the next few decisions at low SNR. Moreover, this compensa-
tion effect is not limited to the maximal distortion channels. This follows because at low SNR
all but one of the tap gains on the forward-TDL are relatively small, hence the sidelobes oc-
curring after the mainlobe still strongly resemble the sampled channel autocorrelation function.
Likewise, the noise at the output still has a correlation function determined nearly by the
matched filter alone, and it also strongly resembles the sampled channel autocorrelation func-
tion. Thus, we observe that this compensation effect will take place at low SNR, regardless
of the particular channel we might care to consider.
Referring again to Figs. 31(a) and (b), we observe that the mean recovery time increases
rapidly beyond about 2 db, for two reasons. The first factor is that the compensation effect
we discussed in the preceding paragraph becomes less as the SNR becomes higher, since grad-
ually the forward-TDL gains increase, causing a mismatched filter whose output noise correla-
tion no longer agrees with the sidelobes occurring after the mainlobe. The second (and more
important) factor, however, is tied in with the properties of the correlation functions of the
output distortions, of the noise and of the intersymbol interference, in the absence of decision
errors, and with an error-propagation effect. We will therefore interrupt our discussion of
the mean recovery time of Figs. 31(a) and (b) to consider these other issues in more detail.
In Figs. 33(a) and (b), we illustrate typical behavior of the output distortion correlation
functions, in the absence of decision errors, by drawing them for the 5-21 case for various
SNR. At each SNR, we have indicated the intersymbol interference distortion correlation func-
tion, the output noise distortion correlation function, and the total output distortion correlation
function (which is the sum of these two). All the correlation functions are shown in normalized
form, with actual values of their distortion energy as indicated. At low SNR, we see that the
total distortion correlation function is the same as that of the noise, since the output noise
66
/ _ _
energy greatly exceeds that of the intersymbol interference distortion. With increasing SNR,
however, the noise distortion continually decreases, while the intersymbol interference attains
a limiting value, for any forward-TDL of finite length. Thus, with increasing SNR, the total
distortion correlation becomes identical with the limiting distortion correlation function of the
intersymbol interference alone, as we may observe from Fig. 33(b). There are two aspects of
these correlation functions we want to bring out at this time. The first concerns the noise dis-
tortion, for, as we see in Figs. 33(a) and (b), it is becoming nearly uncorrelated from one decision
to another as the SNR increases. Our second observation is similar, namely, that the total dis-
tortion is seen to become virtually uncorrelated at intermediate SNR, from about 8 to 20db, as
the intersymbol interference and noise distortion correlation functions tend to cancel one another
out. The fact that the noise becomes nearly uncorrelated at high SNR enabled us to determine
error-rate and burst statistics at high SNR by "inserting an error" to initiate a burst in the
course of our quasi-simulations using our modified error tree algorithm. As mentioned ear-
lier, such an approach is not legitimate at lower values of SNR, where the noise distortion is
seen to be highly correlated from one decision to the next. The fact that the total distortion
tends to become uncorrelated at intermediate SNR enables us to obtain performance data in that
region, by assuming that the total distortion is indeed exactly uncorrelated, without having to
resort to simulations or quasi-simulations, as we will discuss in some detail below.
Returning to our mean recovery time discussion of Figs. 31(a) and (b), we next want to ac-
count for the steep increase in mean recovery time beyond about 2 db which we encountered with
the maximal distortion channels. As we will see in Sec. VI, the value of Dp is always less than
unity beyond about 2 db; thus, the initial error of a burst cannot arise from harmful intersymbol
interference alone, but the additive noise must also be strongly opposed to a correct decision.
As we observed above from Figs. 33(a) and (b), the noise becomes virtually uncorrelated beyond
about 2 db; hence, after a "large" noise sample occurs to initiate the error burst, the noise
tends to resume its typically small values and, thus, error propagation is seen to take place
because of decision errors alone, and virtually without the further assistance of the noise which
initiated it.
To better understand how error propagation can occur, we consider a typical situation (the
5-21 case at 20 db) and show that those future baud message sequences 1, 2'' . . which are likely
to give rise to a decision error, also are prone to exhibit strong error-propagation behavior. In
=
Fig. 34(a), we indicate the decision-feedback equalizer forward-TDL response to a single 40 +1
pulse transmission, denoting each sidelobe by its polarity only, with the mainlobe position located
between the parallel lines. The sidelobes after the mainlobe are canceled out by the feedback-
TDL in the absence of decision errors. The worst possible message sequence is as shown on
the first line of Fig. 34(b), where the k for k < 0 have been replaced by zeros since their effect
is assumed to have been eliminated through error-free decision feedback. The result of weighting
this message sequence by the sidelobes of Fig. 34(a) is shown in the second line of Fig. 34(b),
where all the products are seen to oppose the mainlobe and a correct decision through our choosing
the worst possible message sequence. Suppose a decision error occurs with this sequence, re-
sulting on the next decision in the shifted sequence indicated in the first line of Fig. 34(c). The
leftmost position is now occupied by +1 or -1 with equal probability, which we have indicated with
a question mark. The decision error has caused the feedback-TDL to enhance rather than elim-
inate the q1 sidelobe contribution, as noted by the + sign immediately after the mainlobe position.
67
__1_1_11_1____111____-_---.-II_--_----
------ -·--L---- I
1 3-42-107841
1.200
II1I,
. . --- IlII
IIII !!"1
II Il
LJJI. .... -w db
I.
.'
I..
II I'
rl
._ Idb
*1[
'1
.11I al. ·
-4 db
' 11 - 2 db
' I [I
, .... , . ..., ,1 , .. _
8 db
I' r
.- I -r
(a)
Fig. 33(a-b). Distortion correlation functions for 5-21 case. Left column: intersymbol
interference alone; center column: noise alone; right column: total distortion, sum of
intersymbol interference and noise distortions.
68
0.0244 0.1086 0.1330
II i
'1''I II Ill
I I Ii h
III Il''
-:
I
...... .I 1.......
.- _-.- --
~-.- I. -.| s
I I I I
..... ....
…I .
I I
I I 'wI
, |= . "- 14db
I l, _ nil
-... ·
.....
·.. .. =. Bl
........
I
_& .
Ie ann.J~ .I - II1 I I II I· Onlk
I I I I I ' I " -11 I1' , --
1' L ..
ra
I
.l . . . . .
I I L
~~
~~~ ~-r "
.~"
-- .111 1111 I·u
III
Ii.r i iii
I
i.AiUl
I I
l l
26 db
II III I' I1 IIII II -I
I I
..
11 ,I, I
.. .
II 11
..
I
.
I I I II
IlI .
Ill'
II 1 I I I I I
II 11 1 I II II it 1. 11 I I
I 11 II rT
l l ... .A..-LMAU dL-
- - Is YII[4II. I It. 38db
_ _ -
'I I i 1 I I
(b)
69
(a) + + - + + ++ + + +
+ + +
- - + ++ - - + + +- - + + + + 00
0 0
(b)
+ 0000
?? + + -++ - - + + - + + + -+00
(d)
?? - + + + + - + + + + - + + + + - + + -+00
?? ? + + + - - ++ + _ - +++ -_ + + + -- +0
(e)
?? ? + + - + + + + - + + + + - + + + + + _ _ + 0
Fig. 34. Error propagation example using 5-21 case at 20 db. (a) Equivalent signal
filter output for single +1 input sample; (b) worst possible message sequence for
making decision on t = 1, and their corresponding weightings of sidelobes of (a);
and (c) to (e), shifted versions of this sequence on successive bauds, with errors on
each decision, and with corresponding sidelobe weightings.
The result of weighting this message sequence by the corresponding sidelobes of Fig. 3 4(a) is
shown in the second line of Fig. 34(c). Now, not all the distortions of the intersymbol inter-
ference are opposing the correct decision; in fact, the majority are aiding a correct decision.
Note, however, that our previous decision error has caused an additional sidelobe of 2q, which
being large and positive leads to a decision error on the present baud with probability close to
unity. Proceeding to Fig. 34(d), the first two errors have led to sidelobes - 2q, and 2q 2 , which
are both large, but fortunately tending to cancel one another out. Unfortunately, the majority
of the other sidelobes are now in opposition to a correct decision, thus a third error is likely
to occur with some high probability. Matters continue on in this way, with decreasing but high
probabilities of error on successive bauds, until reaching the point where the first question
mark comes up for a decision. At this point, all the sidelobes occurring before the mainlobe
are being weighted by messages which are independent of the initial message sequence which
helped initiate the burst; thus, the probability of error on each baud once again becomes no
worse than 1/2 and, in fact, it is generally much less than 1/2 due to the relatively low noise
levels in this region. Similar comments are seen to apply to Fig. 34(e). Thus, summarizing,
an initial error caused by the collaboration of harmful intersymbol interference and additive
noise is seen capable of propagating, virtually without further assistance from the additive noise,
due to the particular message sequence which gave rise to the initial error, coupled with the er-
roneous decision feedback. This is what is happening in Figs. 31(a) and (b), where the mean re-
covery time is seen to increase beyond about 2 db for each of the cases studied, eventually reach-
ing maximum values as shown in these two figures. Note that the mean recovery times of Fig. 31(b)
become longer than the times in the corresponding cases of Fig. 31(a), since the number of taps on
the equivalent signal filter is larger in the N-21 than in the N-ll cases, leading to correspond-
ingly longer periods of error propagation such as that encountered in our example above. At this
70
___ __I _
point, we should note that the error-propagation effect leading to increased mean recovery times
will not occur, in general, with an arbitrary choice of a message sequence; hence, if an initial
error were to be caused by noise acting alone (perhaps by impulse noise on a telephone line, for
example), the mean recovery time would generally be considerably shorter than those indicated
in Figs. 31(a) and (b) which were obtained for bursts arising, to a great extent at least, from the
intersymbol interference itself.
Having plotted the mean recovery time and explained its behavior to some extent as a func-
tion of SNR, the most encouraging conclusion we may draw from Figs. 31(a) and (b) is that in all
cases the mean recovery time is bounded, and that on the average no error burst will last indef-
initely. Coupled with the mean recovery time, of course, is the question of recovery time vari-
ance or, correspondingly, its standard deviation. We have plotted the standard deviations of the
mean recovery times for the N-11 and N-21 cases in Figs. 35(a) and (b), respectively.
The general behavior of Figs. 35(a) and (b) is similar to that of Figs. 31(a) and (b) which show
the mean recovery time. In all cases, moreover, we find that the standard deviation of the re-
covery time is less than its mean. Also, due to the longer forward-TDL lengths involved, the
error propagation leads not only to greater mean recovery times in the N-21 than in the N-11
cases as noted previously, but it also results in a slightly greater recovery time variance. The
recovery time standard deviation decreases at very high SNR, since only a small number of side-
lobe combinations can result in a decision error, and since the noise is small, the resulting error
burst is virtually deterministic for the next several succeeding bauds, becoming progressively
less deterministic as new and independent binary signal samples enter into the equivalent signal
filter. We see that the mean recovery time and the recovery time variance are both well-behaved,
thus error propagation will not continue indefinitely with any finite probability. This conclusion
is supported also by our simulation studies, as we illustrate in Figs. 36(a) and (b) which show the
maximum recovery times in bauds which we encountered in our simulations, plotting them vs
SNR for the N-11 and N-21 cases, respectively. From these figures, we note that in all cases
X,
-:1
Z
o
o
a
0 z
a
a
Z
9
W
2
zE
'r0
_.
W
Wt
v,
o
M
0 X
0
cr
0
0
XW
'r
Fig. 35(a-b). Recovery time standard deviation vs SNR for 11- and 21-tap decision-feedback equalizers,
applied to maximal distortion channels.
71
_
1
100
atoo
- I2 50
w
w 50
0 0
U O
20 7- t 20
Z
M.
10
~
____S 22- I
10
_ ,
5 I I I I I I 5 I I I I L - I -
-22 -16 -10 -4 2 8 14 20 -22 -16 -10 -4 2 8 14 20
Fig. 36(a-b). Maximum recovery time vs SNR encountered in simulations of 11- and 21-tap
decision-feedback equalizers for maximal distortion channels.
the maximum recovery times encountered were less than about 5 or 6 times the corresponding
mean recovery times, as shown in Figs. 31(a) and (b), confirming further our conclusion that re-
covery of the decision-feedback equalizer will indeed occur, and that error bursts of indefinite
duration do not pose a real problem in its operation.
Next, we want to consider the guard-space data obtained during our quasi-simulations of
the decision-feedback equalizer. We have plotted the mean guard space as defined in Eq.(161)
in Figs. 37(a) and (b) for the N-11 and N-21 cases, respectively. We observe from these figures
that the mean guard space increases rapidly beyond about 14db. This is to be expected, since
the mean guard space is the reciprocal of the probability of an initial error which, in turn, is
of the same order of magnitude as the overall probability of error, the two being related by a
factor equal to the mean number of errors per burst at high SNR, which is even less than the
mean recovery time. Thus, as asserted earlier in this Section, the mean guard space dominates
the mean recovery time in the denominator of Eq.(159) at high SNR. The large guard spaces en-
countered with the decision-feedback equalizer are very desirable in terms of detecting and cor-
recting the error bursts through coding techniques, a matter we will discuss further in Sec. VI.
We now present a method of obtaining bounds on the mean recovery time of the decision-
feedback equalizer without having to perform simulations, whenever the total output distortion
may be assumed uncorrelated from one decision to another, which we noted earlier was a valid
assumption at intermediate SNR for the maximal distortion channels. Briefly stated, the method
we will develop is to model the error propagation as a discrete Markov chain process, with
"states" corresponding to the contents at the feedback-TDL positions. To obtain the mean re-
covery time exactly, we would have to consider a total of 3 N - 1 such states with an Nth-order
72
-9
(I) -,
'o
I a
0
"C
z
z
LUi w
M,
(/)
Fig. 37(a-b). Mean guard space vs SNR for 11- and 21-tap decision-feedback equalizers,
applied to maximal distortion channels.
channel, since each of the N - 1 feedback-TDL positions may contain no error, a positive error,
or a negative error. Instead, we will ignore the signs of the errors, thus reducing the total
number of states to 2 , with the N - 1 positions either containing an error or not. This leads
us to a means of bounding the mean recovery time by solving the first-passage-time problem in
the associated discrete Markov chain representation of the feedback-TDL state, using appropriate
bounds on the transition probabilities from one state to another. The mathematical formulation
of the problem developed in the following will apply as well to the original set of states, however,
simply by replacing our bounds by the actual transition probabilities.
For an Nth-order channel, the feedback-TDL has N - 1 taps, which we number from right
to left, as shown in Fig. 32. We will represent the state of the feedback-TDL on the nt h deci-
sion by a sequence of digits
where k (n) = 1 if the ith position of the feedback-TDL is occupied by an error on the nt h deci-
sion, and where k(n) = 0 otherwise. We will number these 2 N- states, for convenience as-
N- i
signing numbers 1 and 2 to the error-free and initial-error states, respectively:
The remaining state numbers we assign arbitrarily. Thus, state 1 is the desired operating state
in which no errors are contained in the feedback-TDL, while state 2N - 1 is entered when the initial
error of a burst occurs. After each decision, the contents of the feedback-TDL shift to the left, thus
73
_I-_I--I^--L·LIII11111
__.- · ___UIICIICIII___II·^__-
-^
(n+1) (n)
k i (n+ = k (178)
1 i-I
for i = 2, .. , N-1. The value of kl, of course, depends upon whether or not a decision error
has occurred, thus we see that each state can undergo a transition to exactly two other states.
This is illustrated by the example shown in Fig. 38, where we have indicated the possible transi-
tions for a fifth-order channel. We see there that in the error-free mode of operation the
decision-feedback equalizer occupies the leftmost state shown, the 0000 state. A transition out
of this state to the 0001 state occurs with an initial error. The problem then of determining the
mean recovery time is thus seen to be equivalent to determining the mean first-passage time
from the 0001 state back to the 0000 state in Fig. 38 or, more generally, from state 2 back
to state 1, in view of our assignments (176) and (177). Assuming that an initial error has just
Fig. 38. Error-state transition diagram for Markov chain model of error propagation
in decision-feedback equalizer for fifth-order channel.
occurred, we next solve this mean first-passage-time problem, first introducing some notation
which is generally consistent with that discussed more fully in Ref. 19:
74
____ ___
pij for i 1
ij t0 for i = 1 (183)
From definition of the transition matrix, we have that under these modifications
since initially the feedback-TDL is in state 2 N . If we further define a column vector e 0 with
2 N-1 components via
17'
A
e0 - (187)
0O
then, from notation (182) and Eq.(187), it follows that
which is simply the probability that we are in state 1 after n transitions, conditioned on the fact
that we have not previously entered state 1, through our use of P* instead of P in Eq. (185).
Thus, we conclude that the mean first-passage time Tfp may be expressed by
0oo
nP * n
= [I_--*]-2 p, (191)
n=1
75
N -i 2
where I is the 2 N-1 x 2 identity matrix, simply by multiplying both sides by [I - p,] and
reducing Eq.(191) to an identity. Similarly, noting that the first row of P* contains all zeros
and that the elements in each of the remaining rows sum to unity since some transition must
occur after each decision, then we can verify that
7o-
1
and thus, using Eqs.(191) and (192) in Eq.(190), we find the mean first-passage time is given by
where, again, 7r(0), P*', and e are as defined in Eqs. (186), (184), and (192), respectively.
Next, we want to apply Eq. (193) in order to obtain a bound on the mean recovery time of
the decision-feedback equalizer. To do this, we must specify the appropriate bounds on the
Pi', the elements of the modified transition matrix P*'. There are a number of ways in which
ij
we might do this. We will describe one method, convenient for its simplicity, which has proved
useful in practice. First we observe that the net sum of the uncanceled sidelobes is either tending
to assist or oppose a correct decision. Clearly then, we may determine an upper bound to the
probability of error by first defining for each state a quantity Q via
the summation being over the uncanceled sidelobes, and the factor of two arising from the fact
that these sidelobes were doubled rather than canceled through the erroneous decision feedback.
Then, since under our assumptions the two situations of Eq. (194) are equally probable, we may
write down a bound on the probability of error while in any state:
where De and Q, of course, are dependent upon i, the number of the state being considered.
The calculation of the quantities p(errorlQ) appearing in Eq.(196) is a straightforward applica-
tion of our modified error tree algorithm, being identical with our discussion of finding the
=
probability of the initial error of a burst, with the exception that the mainlobe q 1 used there
76
must now be replaced by a mainlobe of q = + Q. This renders the desired bounds on the
probability of error in each of the feedback-TDL states which, in turn, render the p' needed
in Eq. (193) to determine the bound we desire on the mean recovery time.
The approach indicated in the preceding discussion can thus enable us to arrive at bounds
on the mean recovery time of any channel, of arbitrary order, without having to resort to the
time-consuming methods of simulations or quasi-simulations.
In this Section, we have presented the techniques and algorithms developed for determining
the operation and performance of the decision-feedback equalizer at all SNR, and the results of
applying these methods using the maximal distortion channels as examples were shown and dis-
cussed. Much of the discussion centered on the problem of error bursts, and the properties of
the recovery time. Our final summaries regarding the operation of the decision-feedback equal-
izer are included in Sec. VI, where we also compare it with the operation of the conventional
equalizer studied in Sec. IV.
77
_ _ ___
VI. CONCLUSIONS - SUGGESTIONS FOR FURTHER RESEARCH
79
__I slp· _I
· l
SNR (db)
-16 -10 -4 2 8 14 20 26
SNR (db)
80
1
We may obtain some feeling as to the degree to which the conventional and decision-feedback
equalizers are sub-optimal, through comparing the N = 2 through N = 5 curves of Figs. 39 and 40
with the nondispersive N = 1 curve, since this curve represents an upper bound on the performance
of the optimal equalizer derived in Sec. II. We observe that the conventional equalizer N = 2
-4
curve is about 9.5 db poorer than the N = bound at a probability of error of 0 , with the loss
being even greater for the higher-order maximal distortion channels. On the other hand, from
Fig. 40 for the decision-feedback equalizer in this N = 2 case, the loss in SNR is only about
2.8 db compared with the N = 1 bound, and moreover the loss increases by less than an additional
2 db for the higher-order maximal distortion channels considered, at this same error rate of
10 . Thus, we conclude that, although their hardware-implementation requirements were made
identical, the conventional equalizer is much more sub-optimal than the decision-feedback equal-
izer which, in turn, is so close to the upper bound on performance that the maximum improve-
ment in SNR which could possibly be achieved through any alternative sub-optimal equalization
scheme is probably not worth the additional complexities it would require.
As mentioned in our concluding discussions of Sec. III, it is in the different treatment given
by the decision-feedback equalizer to the intersymbol interference contributions arising from
past and future bauds which accounts for the performance advantages it enjoys over the conven-
tional equalizer. The respective sidelobe-suppression behavior of the two equalizers is illus-
trated in Fig. 41, where we have plotted Da and D vs SNR for the 5-21 case. These curves,
typical of those obtained in all other cases studied, explain why the conventional equalizer per-
formances shown in Figs. 21(a) to (d) reach limiting values with increasing SNR, while the cor-
responding decision-feedback equalizer performances
4.0
shown in Figs. 27(a) to (d) continue to improve with
SNR. At high SNR, the limiting Dp for the conven-
tional equalizer is seen to approach a value of ap-
3.2 proximately 1.81. Thus, if a sufficient number of
sidelobes collaborate to oppose a correct decision,
the net intersymbol interference distortion can ex-
z ceed unity, thus leading to a decision error even in
2.4
the absence of noise. Of course, regardless of Dp,
the probability that this will occur will drop off rapidly
with the overall number of sidelobes; hence, by using
a sufficiently long TDL in the conventional equalizer,
2 1.6
we may obtain any desired performance level at high
SNR. The SNR levels and TDL lengths involved,
however, are much larger than those required using
0.8
the decision-feedback equalizer to obtain the same
performance levels. Of course, this is because
generally the decision-feedback equalizer is able to
devote many more of its total number of taps to re-
0 ducing the intersymbol interference arising from
-16 -4 8 20 32 44
81
.--
decision-feedback equalizer is less than unity beyond approximately 2 db; thus, above this SNR,
the intersymbol interference acting alone cannot cause a decision error, and therefore the per-
formance will continue to improve with SNR, as we observed earlier in Sec. V. To confirm the
sidelobe-suppressing advantage of the decision-feedback equalizer still more, we have plotted
the limiting value of Dp with increasing SNR vs the TDL length of the conventional equalizer in
Fig. 42. Comparison of this figure with Fig. 30 illustrates the higher residual sidelobe distortion
levels obtained with the conventional equalizer, and accounts for its much poorer performance
than that of the decision-feedback equalizer at the high SNR of practical importance.
A further advantage of the decision-feedback equalizer is in terms of its noise enhancement,
which is less than that of the conventional equalizer, as we see in Fig. 43 where their noise
enhancements are plotted vs SNR for the 2-3 case. Despite the fact that the noise enhancement
is smaller with the decision-feedback than with the conventional equalizer, we found above that
it nonetheless accounts for nearly all the degradation in performance encountered with decision-
feedback equalizers of long length at high SNR, whereas with the conventional equalizer the
residual sidelobes were seen to determine its limiting performance, independent of the additive
noise enhancement.
From our recovery-time and guard-space studies of Sec. V, it is clear that at all those error
rates and SNR of practical interest, the mean guard space greatly exceeds the mean burst duration,
and even the maximum burst durations encountered in our simulations. Such large error-free
guard spaces between error bursts are very desirable from the viewpoint of applying coding
techniques to further improve the overall performance of the decision-feedback equalizer. In
the error-detecting and -correcting scheme of Ref. 20, for example, the guard space need only
be three times as long as the longest possible burst duration in order to insure 100-percent
error correction. Such a scheme applied here with the decision-feedback equalizer would have
28
3-2-10198
24
o
z CONVENTIONAL
o ' 20 EOUALIZER
2
en
o0 1
I
v z 16
Z
z z
I- I
2
z
58
z
< 4
12
-
M
Z
w
w DECiSiON-FEEDBACK
V,
EQUALIZER
o a
0 Io I I I I
-16 -10 -4 2 8 14 20 26
TDL LENGTH SNR (db)
Fig. 42. Limiting distortion Dp vs TDL length of Fig. 43. Example of noise enhancement vs SNR
conventional equalizer, applied to maximal dis- for conventional and decision-feedback equalizers
tortion channels. in 2-3 case.
82
___
enabled virtually error-free data transmission. Other schemes requiring even smaller guard
spaces may be employed as well, for although these permit certain error sequences to remain
undetected, such sequences have sufficiently low probability of occurrence that again one attains
virtually error-free data transmission. The method of Ref. 20 is being applied in the construc-
tion of equipment capable of correcting error bursts of up to 1000 bauds in duration, and with
relative simplicity. Thus, we conclude that in view of the error-burst and guard-space data
we obtained on the decision-feedback equalizer, its use in conjunction with coding techniques
for eliminating its error bursts is both quite feasible and highly desirable.
The present work may easily be extended to include the colored noise case, the only essential
difference in structures of the equalizers considered being that the matched filter becomes re-
placed by a "whitening filter" in cascade with a second filter, the latter being matched to the
equivalent channel in cascade with the whitening filter. Also, if the output noise statistics are
non-Gaussian but known, then our algorithms for determining the equalizer performances still
apply, except now the erfc( ) function of Eq. (123) must be replaced throughout the calculations,
and the noise samples given the appropriate statistical distribution in the simulation studies.
Correlated message sequences are easily handled by redefining the total output distortions
of Eqs. (34) and (90) to include the effects of the correlation between the signals upon the inter-
symbol interference component of the distortion. This, however, merely modifies the definitions
of the elements of the X matrices of Eqs. (39) and (82), but otherwise leaves unaltered the tap-
gain vector solutions of both the conventional and decision-feedback equalizers.
Another issue for further investigation is the use of multilevel signaling schemes. Four-
and eight-level digital communication systems are common, and even more levels are used in
some applications. The algorithms used here could be directly extended to such multilevel
systems. The error rate of the decision-feedback equalizer at the higher bit rates attained
through multilevel signaling would, of course, be somewhat poorer than in the binary signaling
scheme considered throughout our studies in this report. This could occur for two reasons.
First, a large noise can cause an error regardless of its polarity in a multilevel signaling
system, whereas with a binary scheme it will always assist a decision with probability of 1/2.
Second, this same effect applies to the large intersymbol interference distortions arising through
erroneous decision feedback, hence we would expect that the recovery times might be somewhat
longer than with a binary signaling system. The degree to which these factors are important
thus should also prove a matter of interest in future research.
The decision-feedback equalizer structure, as derived under our assumptions in Sec. III,
involves a nonlinear feedback loop, inasmuch as we may consider the decision-feedback opera-
tion to be implemented by our first passing the output samples through a hard limiter, and then
taking the limiter output as the feedback -TDL input as shown in Fig. 4 4(a). Alternative schemes
which might prove even more effective readily come to mind. For example, if the output statistic
were passed through the "soft-limiter" of Fig. 44(b), an improvement in performance might be
realized. Looking into different types of nonlinear feedback such as this is thus another topic
of interest.
Finally, and perhaps one of the most important issues remaining to be studied, is the
question of the sensitivity of the performances of the conventional and decision-feedback equal-
izers to channel measurement errors. In all the studies presented here, of course, perfect
channel measurement was assumed; as a result, the tap gains of both equalizers were able to
83
---__I1_II l-_-·I_^IXI^-_C·l-_I-----
---
-
13-42 -1019
(0)
/ )I
DECISION
STATISTIC OUTPUT TO
INPUT FEEDBACK -TDL
I it
(b) \-
I I
be set optimally under our output distortion measure. With channel measurement errors, how-
ever, the equalizer performances must deteriorate. The question of sensitivity is especially
of interest in the case of the decision-feedback equalizer, which allows large sidelobes to occur
after the mainlobe (recall Fig. 10) with the expectation that they will be canceled out by the
feedback-TDL output. With channel measurement errors, even in the absence of previous
decision errors, such cancellation will be imperfect, leading to poorer performance. Thus,
also requiring investigation is the extent to which such measurement errors affect the high
performance levels achieved by the decision-feedback equalizer in our studies, when compared
with the type of conventional equalizer currently employed.
ACKNOWLEDGMENTS
84
REFERENCES
85
.-.--
---.-·-II-. . __LI____IIII__··___·I- IICI 1-· --
JOINT SERVICES ELECTRONICS PROGRAM
REPORTS DISTRIBUTION LIST
U
JOINT SERVICES REPORTS DISTRIBUTION LIST (continued)
II ______·__IIYU___^_ ---- · ·
-----
LIY-·III(·I-L.tL _-tlll·ll-- I_ ·- X _--
JOINT SERVICES REPORTS DISTRIBUTION LIST (continued)
Non-Government Agencies
Other Government Agencies
Director
Mr. Charles F. Yost Research Laboratory of Electronics
Special Assistant to the Director Massachusetts Institute of Technology
of Research Cambridge, Massachusetts 02139
National Aeronautics and
Space Administration
Polytechnic Institute of Brooklyn
Washington, D.C. 20546
55 Johnson Street
Brooklyn, New York 11201
Dr. H. Harrison, Code RRE Attn: Mr. Jerome Fox
Chief, Electrophysics Branch Research Coordinator
National Aeronautics and
Space Administration
Director
Washington, D.C. 20546
Columbia Radiation Laboratory
Columbia University
Goddard Space Flight Center 538 West 120th Street
National Aeronautics and New York, New York 10027
Space Administration
Attn: Library C3/TDL Director
Green Belt, Maryland 20771 Coordinated Science Laboratory
University of Illinois
NASA Lewis Research Center Urbana, Illinois 61803
Attn: Library
21000 Brookpark Road Director
Cleveland, Ohio 44135 Stanford Electronics Laboratories
Stanford University
Stanford, California 94305
National Science Foundation
Attn: Dr. John R. Lehmann Director
Division of Engineering Electronics Research Laboratory
1800 G Street, N.W. University of California
Washington, D.C. 20550 Berkeley, California 94720
Director
Microwave Laboratory
Stanford University
Stanford, California 94305
Technical Report
5. AU THOR(S) (Firstname, middle initial, last name)
Michael E. Austin
c. (2) Contract N00 140- 67-C-0201 Sb. OTHER REPORT NO(S) (Any other numbers that may be assigned
this report)
ESD-TR- 67-466; Lincoln Laboratory
Technical Report 437
10. DISTRIBUTION STATEMENT
This document has been approved for public release and sale; its distribution is
unlimited.
II. SUPPLEMENTARY NOTES 12. SPONSORING MILITARY ACTIVITY
Joint Services Electronics Electronics
Program thru USAECOM, Fort Monmouti
N. J. and U. S. Navy Purchasing Office
nis
I. report, a decision-theory approach to the problem of digital communication
over known dispersive channels is adopted to arrive at the structures of the optimal and
two sub-optimal receivers: the "conventional" and "decision-feedback" equalizers, both
employing matched and transversal filters. The conventional equalizer is similar to
some equalization modems finding current application, while the new decision-feedback
equalizer (as its name implies) utilizes its previous decisions in making a decision on
the present baud.
The parameters of these two equalizers are optimized under a criterion which mini-
mizes the sum of the intersymbol interference and additive noise distortions appearing
at their outputs. Algorithms for evaluating the performances of the conventional and
decision-feedback equalizers are developed, proving especially useful at high SNR
where simulation techniques become ineffective.
To determine the dependence of their performance and sidelobe-suppression prop-
erties upon SNR and transversal filter length, the new algorithms were used to study the
conventional and decision-feedback equalizers when they are applied in equalization
of the class of channels exhibiting the maximum realizable intersymbol interference.
In addition, error-propagation effects found to arise in the decision-feedback equalize:
operation are studied. High overall error rates (as found at low SNR) lead to a threshold
effect as the decision-feedback equalizer performance becomes poorer than that of the
conventional equalizer.
Despite such error-propagation behavior, however, the decision-feedback equalizer
is found to be considerably better than the conventional equalizer at all SNR and error
1-
rates of practical importance, due to its sidelobe-suppression behavior; moreover, its
'. . advantages become more pronounced, the higher the SNR and the greater the channel
dispersion.
DD FORM6 5 1 4 7 3 DDSN
(PAGE~ 1)
00-0-6~13
NOV
UNSurtyCUNl,A
ssiTFiaTi)
S/N 0101-807-6811 Security Classification A-31408
A- 31408
---
___i_~_~'--~
-"--- - -- -~
UNCLASSIFIED
Security Classification
.
14.
LINK A LINK B LINK C
KEY WORDS
ROLE WT ROLE WT ROLE WT
_-w
decision-feedback
digital communication
dispersive channels
equalization filters
threshold effects
I
rI
I ii I
p .... imm r _W
m,, I
-----·--- -- ----